Artemis release notes - Kore.ai Docs

v1.8.3

This release expands the platform’s capabilities in agent building, voice, digital channels, and governance. Key features include language detection for customer messages, WhatsApp invitations during voice calls, configurable supervisor return thresholds, extra Genesys transfer headers, and Bitbucket Data Center support on private networks.

Agent builder experience

Detect the language of each customer message

Projects can now detect the language of every customer message and share the result with agents as session.languageDetection, including the language, confidence, and status. Turn it on in Runtime Config under the new Language detection section, where you can also set a confidence threshold, a minimum message length, and the allowed languages. Detection doesn’t change the conversation language on its own. An agent’s before_turn hook decides, using WHEN conditions with SET_INTERACTION_LANGUAGE or CLEAR_INTERACTION_LANGUAGE. The reply, the filler, and, on calls without a voice language profile, the voice then switch together. Session traces show the detection result under each customer message. The setting is off by default and can be turned on once the language detection service is available in your environment. Learn more →

HOOKS:
  before_turn:
    WHEN: session.languageDetection.status == "detected" AND session.languageDetection.confidence >= 0.85
      SET_INTERACTION_LANGUAGE: session.languageDetection.language

# Switches the reply language only when detection is confident. A low-confidence result leaves the language unchanged.

Control when callers return to the supervisor

While a specialist is handling a conversation, the supervisor’s classifier checks whether the caller’s reply belongs to a different specialist. You can now set how confident it must be before the caller goes back to the supervisor. In ‘Runtime Config > Reasoning pipeline’, turn on ‘Custom Return Threshold’ and set the threshold. To leave unclear replies with the current specialist, turn on ‘Keep Ambiguous Replies With Specialist’. A supervisor can override both settings in its EXECUTION pipeline with parentReturn. The classifier also now judges the reply against the specialist’s own conversation, so callers are returned by mistake less often. Learn more →

EXECUTION:
  pipeline:
    enabled: true
    parentReturn:
      minConfidence: 0.9
      ignoreAmbiguous: true

# The caller returns to this supervisor only when the classifier is at least 90% sure, and never on an ambiguous result.

Run session-end workflows for several agents

The After Session Ends workflow trigger can now target more than one agent. Select the agents in the trigger’s Agents list. The workflow runs when a session owned by any of them ends, but not for sessions handled only by internal helper agents. Leave the list empty to run the workflow for any agent. Existing triggers keep their selected agent. Learn more →

Voice experience

Send continuity invitations over WhatsApp during a call

During a voice call, an agent can now send the caller a WhatsApp message with tappable options, such as appointment slots, using the offer_continuity tool with delivery: 'whatsapp'. The message uses a WhatsApp template and goes through the project’s WhatsApp connection on Meta Cloud or Infobip. The caller’s tap or typed reply reaches the live call as input, on both pipeline and realtime voice. Replies that arrive after the invitation expires don’t reach the call, and the caller gets a message saying the request expired. Text typed on a web page that has joined a live pipeline voice call is also answered on the call.

More transfer headers for Genesys Audio Connector

Escalations to Genesys through the Genesys Audio Connector can now send any extra headers your Genesys flow needs. Add them under additional_headers in genesys_sip_headers, and each one is sent as its own header with the transfer. Values can use {{...}} templates that are filled in from the session at transfer time, and a trigger’s headers override the routing defaults one by one. On korevg_savg transfers, the SIP BYE now also carries User-to-User Information (UUI), and transfer metadata is no longer limited in size. Learn more →

ESCALATE:
  routing:
    connection: "genesys_prod"
    genesys_sip_headers:
      additional_headers:
        X-Interaction-Id: "{{interactionID}}"
        X-Brand: "{{brand}}"

# Each additional header is sent separately with the transfer, with its template filled in at transfer time.

Digital experience

Let Genesys Architect announce escalations

For Genesys Cloud Bot Connector, you can now leave the escalation message to Genesys Architect, such as to check business hours before telling the customer about the transfer. Set escalation_format to an empty string in the agent’s MESSAGES. A successful escalation then returns no reply messages, while still sending the escalation intent. The same applies when the response is replayed. If you leave the key out or set a message, escalations work as before. Learn more →

MESSAGES:
  escalation_format: ""

# Genesys Architect plays its own transfer announcement. The escalation intent is still returned.

Connect Infobip through an API gateway

Infobip WhatsApp connections can now authenticate with a custom header, for setups that route Infobip traffic through an API gateway such as Azure API Management. Choose Custom Header (API Gateway) as the Auth Type and enter the header name and value. The connection dialog now shows only the fields for the selected auth type. You can also add an optional inbound webhook secret, which lets the connection accept typed replies to WhatsApp continuity invitations. Learn more →

Governance

Bitbucket Data Center on private networks and through your proxy

Projects can now sync with a Bitbucket Data Center server on a private network, such as one you reach over VPN or a private link. Allow the host on the Basic auth profile used for the repository, and it can resolve to a private address for that profile’s repositories. If the server is behind a corporate proxy, add a Global rule for it in Settings > Account > Secrets > Proxy Configs. Git and API calls to the server then go through that proxy. When Studio can’t resolve the host, the error now says so, instead of reporting that the host isn’t allowed.

v1.8.2

This release gives you finer control over reasoning steps, including limits on which session fields the model can write, and lets you reuse MCP connections within a conversation. Voice updates include gateway turn detection, softphone links, and ElevenLabs v3/v4 models. Digital channels add Amazon Connect live-agent transfer and password fields in action sets, while Governance brings mutual TLS for MCP servers.

Agent builder experience

Reuse MCP connections within a conversation

MCP servers have a new Reuse connections setting. When it’s on, the agent keeps one connection to the server open for the whole conversation instead of reconnecting for every tool call, which makes tool calls faster, especially on voice. Workflows that an agent starts share the agent’s connection, so stateful MCP servers keep their state across those calls. The setting is on by default for new servers and off for existing ones.

Choose how much specialists see of other requests

When governed multi-intent handling is on, the new ‘Request view for specialists’ setting in Runtime Config controls what the specialist serving a request can do with the customer’s other requests.

Off: Shows only the specialist’s own conversation.
Inspect: Shows the other waiting requests and lets the specialist suggest a new one.
Manage: Also lets the specialist cancel or change a waiting request, but only when the customer asks for that change in their own words in the same turn.

Limit which session fields a reasoning step can write

A reasoning step’s CONTEXT_UPDATES setting can now list the exact session fields the model is allowed to write, instead of only turning writes on or off. The model can update only the listed fields, and an empty list ([]) blocks all model-supplied updates in that step. Authored actions such as SET, GATHER extraction, hooks, and tool result mappings work as before. To use lists, every Runtime instance must run v1.8.2 before you publish the agent. Existing agents keep their current behavior until you change the step.
Learn more →

FLOW:
  write_qr_message:
    REASONING: true
    GOAL: "Draft the QR code message and mark it done."
    CONTEXT_UPDATES: [qr_text, qr_done]
    EXIT_WHEN: qr_done == true

# The model can write only qr_text and qr_done in this step. Every listed field must be declared in MEMORY.session.

Hand back to the parent agent after a one-way handoff

A handoff with EXPECT_RETURN: false still doesn’t return automatically. However, the receiving agent can now choose an authorized HANDOFF route back to the agent that called it during the same conversation, such as after the customer changes topic.

Voice experience

Voice gateway turn detection

Pipeline voice channels can now let the Kore.ai Voice Gateway decide when the caller has finished speaking, using voice and wording signals instead of a fixed silence timer. On the channel’s speech settings, turn on ‘Voice gateway turn detection’ and choose a Patience level, from Fastest (0.3) to Slow dictation (0.9). Faster levels reply sooner, while the Dictation and Slow dictation levels wait through long pauses when callers spell names or read out card and reference numbers. You can also set turn detection for an agent, a behavior profile, a response, a flow step, or a single GATHER field, so only the questions that need more patience wait longer. Learn more →

FLOW:
  get_card_number:
    REASONING: false
    VOICE:
      turn_detection: { enabled: true, eou_threshold: 0.8 }
    GATHER:
      - card_number:
          TYPE: string
          PROMPT: "Please read out your 16-digit card number."

# Dictation patience (0.8) applies only while this step asks for the card number. Other steps keep the channel setting.

Listen for words found earlier in the call

A GATHER field’s speech vocabulary can now come from a session value set earlier in the call, not just from fixed words in the agent. For example, after a lookup tool returns the streets in the caller’s county, the street question listens for those street names only while that question is asked. Values that hold personal data are sent to the speech recognizer only when the field allows it, and each use is recorded in the PII audit log.

Interruption, silence and transcript controls for OpenAI Realtime

OpenAI Realtime voice channels have new controls under Advanced voice behavior.

Caller interruption: Decides whether callers can cut off the agent and protects messages that shouldn’t be interrupted, such as the welcome.
Silent caller reprompt: Applies the project’s no-input settings, including timeout, reprompts, and final action, to realtime calls.
Caller transcript cross-check: Gives the agent the caller’s transcript alongside what the model heard, catching dropped digits and misspellings.
Learn more →

Shareable softphone links

You can now share a voice channel with anyone through a link. On the channel’s Testing tab, generate a link and send it to a tester or stakeholder. They can call the agent from a phone browser through the real voice gateway and follow a live, read-only transcript. Each link is shown only once and expires after 7 days. You can view and revoke active links on the same tab.

ElevenLabs v3 and v4 dialogue models

The ElevenLabs TTS Model list for voice channels now includes Eleven v3, Eleven v4, and Eleven v4 Turbo, at both the channel level and per language. Calls now use the model you select instead of switching to Flash v2.5.

Pause between a filler and the answer

Pipeline voice channels have a new Filler to answer pause setting in Runtime Config. It adds a short pause (0 to 5,000 ms) after a filler phrase finishes and before the agent speaks the answer, so the two don’t run together. Voice latency profiles inherit the value, and you can override it per profile. The default is 0, so current calls aren’t affected.

Save spoken fillers in the conversation history

A new ‘Include spoken fillers in conversation history’ setting in Runtime Config saves each spoken filler, such as “One moment while I check that,” as an agent message. Fillers are saved only after they finish playing, so transcripts match what the caller heard. The setting is off by default.

More reliable Bring Your Own SIP setup

Saving a Bring Your Own SIP voice channel now keeps your configuration, even if the voice gateway can’t be updated right away. Studio shows a message, and you can select Save again once the gateway recovers. Entering a phone number (DID) now enables call routing, and clearing a number removes its mapping. The SIP gateway field accepts up to ten IP addresses or CIDR ranges.

Digital experience

Password fields in action sets

Action set input fields can now use INPUT_TYPE: password for PINs, passcodes, and passwords. Web chat masks the value as the customer types. When PII protection is on, a protected token replaces the submitted value before it reaches traces, history, or session variables, while tools still receive the original value. Action set elements also now appear in the order you declare them, so a button placed after the input fields stays after them. Learn more →

ACTIONS:
  - id: phone_pin
    type: input
    label: "PIN"
    inputType: password
    required: true
  submitLabel: "Sign in"
  submitId: "submit_login"

# The PIN is masked as the customer types it, and protected in traces and history when PII protection is on.

Transfer chats to Amazon Connect agents

Amazon Connect is now available as an Agent Desktop provider for live-agent transfer. Create an Amazon Connect connection with your AWS Region, Connect instance, and contact flow, then route an ESCALATE to it. The platform starts the Connect chat for the customer, who stays in the same chat window while the agent replies from the Amazon Connect Contact Control Panel (CCP). AWS credentials stay on the server, and no customer-hosted code is needed.

Same-channel conversation recall is now opt-in

Cross-channel recall no longer brings earlier conversations from the channel in use into a new session. For example, a new phone call no longer recalls the customer’s earlier phone calls. To include them, turn on ‘Include earlier conversations on the same channel’ under Settings > Omnichannel > Conversation Recall. The agent now sees recalled conversations in the background only and always asks again for identity and verification details.

Simpler Microsoft Agent 365 channel setup

Microsoft Agent 365 connections are now managed entirely from the channel’s Configuration tab, which shows the tenant ID and blueprint ID and lets you rotate the blueprint secret. The separate Credentials tab, including the Use Auth Profile option that could break replies, no longer appears for this channel. The tenant ID and blueprint ID can’t be changed after creation. To use different values, create a new connection. The Agent 365 tile now appears only once on the Channels page.

Governance

Mutual TLS for MCP servers

MCP servers can now present a client certificate when they connect. In Project Tools > MCP Servers, open Advanced Settings and select an mTLS profile. It works separately from the server’s authentication profile, and both apply to every tool on the server. If a selected profile is disabled or can’t be used, the connection is refused with a clear error, instead of attempting the connection without it.

View the SIP Ladder with session access

Anyone who can open a session can now view its SIP Ladder on the Voice tab, with call endpoints redacted. This view no longer needs the separate SIP Ladder reveal permission. Revealing personal data and raw provider payloads still requires its own permissions.

v1.8.1

This release introduces Bitbucket Data Center support for Git integration and session-aware behavior profiles that activate based on runtime context, such as customer tier. Voice enhancements include turn fillers for OpenAI Realtime, dedicated speech recognition for caller transcripts, and per-field keypad settings. Governance now supports configurable guardrail coverage modes that limit compliance checks to customer-facing interactions.

Agent builder experience

Session-aware behavior profiles

Behavior profile WHEN conditions can now access approved runtime and session variables, including declared session variables, GATHER fields, tool results, input, and intent. This lets profiles activate based on business state, such as once the customer’s tier is known. Learn more →

GATHER:
  customer_tier:
    prompt: "What is your membership tier?"
    type: enum
    options: [standard, gold, platinum]

BEHAVIOR_PROFILE: vip-handling
PRIORITY: 10
WHEN: customer_tier == "gold" OR customer_tier == "platinum"
INSTRUCTIONS: "Greet the customer by name. Offer the priority support line before closing."

# Once customer_tier is collected (by GATHER, a tool result, or SET), the profile activates on the next turn, no context.* prefix needed for an ordinary session value.

Clearer handoff timing and failures in session traces

For handoff and delegate steps, session traces now show the target agent’s run time separately from the routing overhead. A handoff no longer looks slow just because the next agent took time to respond. Refused handoffs now also appear in the trace and in the session’s error list, along with the target agent and the reason.

Knowledge base

Save all field mappings at once

The connector field-mapping review now has a Save all mappings action that saves the mappings for every object type in one step. Mappings also stay more consistent across syncs, and SharePoint sources migrated to the packaged connector now handle their sync lifecycle correctly.

Voice experience

Turn fillers for OpenAI Realtime channels

OpenAI Realtime voice channels can now play short hold phrases while the agent works on a turn. You can turn on the ‘Realtime turn filler’ on the channel and, optionally, override the filler prompt. The model words each filler, and the agent’s answer waits until the filler finishes, so they never talk over each other.
Learn more →

Dedicated speech recognition for caller transcripts

Realtime LLM voice channels can now run a dedicated speech-to-text provider alongside the speech-to-speech model. Session transcripts, saved messages, and traces then record what the caller said more accurately, while the realtime model still runs the conversation. A platform administrator enables this per workspace. In the channel’s Caller transcription card, you can choose the provider and language. Also, Gemini (STT) is now available as a provider, with end-of-speech tuning, and the session’s voice metrics show which source each caller transcript came from.

Keypad input settings for GATHER fields

A voice flow GATHER field can now include a DTMF: block that sets the digit count, finish key, and pause between key presses for that field. On Kore Voice Gateway calls, these settings apply while the field is being asked, and the call’s normal keypad settings return afterward. The compiler also warns when a GATHER field setting won’t be read, such as an unknown key or a multi-line VALIDATE. Learn more →

FLOW:
  get_mobile:
    REASONING: false
    GATHER:
      - mobile_number: required
        type: string
        prompt: "Please say or enter your 10-digit registered mobile number."
        DTMF: { digits: 1..10, finishOnKey: "#", interDigitTimeoutMs: 5000 }
        VALIDATE: "^[0-9]{10}$"
    THEN: verify_customer

  # The caller can type up to 10 digits, press # to finish early, or pause 5 seconds to end entry. Can still just say the number instead. 

Governance

Bitbucket Data Center for Git integration

Projects can now sync with repositories hosted on Bitbucket Data Center, not just Bitbucket Cloud. Connect with an HTTPS repository URL and a Basic auth profile that holds a personal HTTP access token. Then allow the Data Center host to receive that profile’s credentials. Studio either registers the push webhook for you or shows the URL and secret, so a repository admin can add it. For production branches, you can require a project-owned auth profile, so tenant-wide profiles can’t be used.

Guardrail coverage modes

Project guardrail policies have a new Coverage setting. It controls which text a policy checks.

All model I/O (default): Checks every model exchange, including tool inputs, tool results, and sub-agent handoffs. Behavior is unchanged.
User boundary: Checks only what the customer writes and the responses the customer sees. Use it when internal exchanges don’t need policy checks, and you want fewer checks per turn.

Studio shows a banner and a save warning that list the rules that stop running. Agent-defined guardrails and untrusted-text scanning always run.
Learn more →

v1.8.0

This release expands the platform’s capabilities in agent building, voice, knowledge base, LLM models, and governance. Key features include Decision Models for intelligent request routing, new model support across Claude 5, GPT-6, and Gemini 3.x, business rules for knowledge base search, Grok and Microsoft Speech voice provider options, PostgreSQL, Oracle DB, and Microsoft Exchange on-premises connectors, and JWT App credentials with expanded PII redaction coverage.

Agent builder experience

Decision Models for faster, cheaper routing, and conversation checks

Decision Models add a small, fast model that answers quick checks, such as which specialist should handle a request or whether it’s one request or two, before the full AI model does. It responds in under half a second, compared with the AI model’s 1 to 2 seconds, so confident answers skip the full model call and cut latency and cost. Uncertain decisions still go to the full AI model, as they do today. Admins can turn it on, pick a provider, and each project can opt in per step.

PostgreSQL connector

Agents and workflows can now connect to PostgreSQL databases. Supported actions include Get Tables, Get Table Schema, Find Rows (with column filters, ordering, and pagination of up to 500 rows), Insert Row, Update Rows, Delete Rows, and Run Query with parameterized binds. Update and Delete operations refuse an empty match to prevent accidental bulk changes.

Microsoft Exchange on-premises connector (EWS)

Agents and workflows can now connect to on-premises Microsoft Exchange Server mailboxes through Exchange Web Services (EWS). Supported actions include sending, finding, and reading emails, retrieving attachments, and running custom SOAP requests. A polling trigger watches for new emails. This is separate from the existing Microsoft Outlook connector, which reaches Exchange Online through Microsoft Graph.

Autoloop approval modes

Autoloop now supports three approval modes: Autopilot applies changes automatically, Copilot applies them after your approval, and Advisor suggests changes only. You can set the mode when starting a loop or change it mid-run, and it carries over when you restart a loop or switch the source evaluation. The server also automatically advances stalled loops.

Router auth profiles and credential testing

Auth profiles now support router mode (MODE: router), which selects another auth profile for each tool call based on parameter values, such as brand or country. Parameters used for routing can be marked as Auth only, so they aren’t sent to the API. Router profiles don’t store credentials themselves. Test Credentials now supports request-signing profiles, including SigV4 and DPoP, and shows which profile a router selects. Guard conditions can use ABL expressions to route requests based on response status and other supported context.

Guardrail access to tool arguments and results

Output guardrail checks can now reference the pending tool call’s arguments, this turn’s tool results, and the full session state in their conditions.
Learn more →

Hand off to a specialist in ON_START

ON_START now accepts a DO list that includes HANDOFF, so an entry agent can route the session to a specialist before the first user message arrives. This is useful for inbound voice calls where the channel or caller data determines which agent should handle the call. After the specialist returns or fails, the entry agent resumes with its own model settings.

Output schemas for MCP tools

Studio now displays the output schema of an MCP tool. At runtime, when an MCP tool runs, the platform also passes the tool’s output schema to the model along with the tool response. This gives the model context about the output parameters, so it can handle the response correctly.

Access raw speech transcripts in flow steps

A new read-only input_transcript value gives flow steps, hooks, and reasoning logic the speech provider’s raw transcription of the current caller turn. This lets voice agents access what the caller said, without relying solely on the model’s interpretation.
Learn more →

Control which memory entries the model sees

You can now mark individual automatic memory entries as hidden from the model. Hidden entries remain available in expressions and tool calls but are excluded from the model’s prompt, so sensitive or noisy context does not influence the model’s reasoning. Set visibility per memory field in the agent editor or in ABL.

Set interaction language in ON_START

ON_START handlers now accept SET_INTERACTION_LANGUAGE and CLEAR_INTERACTION_LANGUAGE. You can set the conversation language during startup, so the first response already uses the correct language without waiting for the caller to speak. This is especially useful for voice agents whose channel or SIP headers indicate the caller’s language.

Module agents limited to declared tools

An agent imported from a reusable module now receives only the tools declared in their own definition, not every tool the module exports. This prevents an agent from seeing or calling tools its author never assigned to it, matching standalone agents’ behavior.

Credential header validation for HTTP tools

The compiler now warns when an HTTP tool’s headers contain a credential-shaped value, such as an Authorization or API-key header written as plain text instead of a secret reference. Studio shows these diagnostics on the tool detail and list pages, letting you find and fix exposed credentials before deploying. This is a warning, not a blocking error, so an existing tool with an inline credential still compiles and runs.

Redesigned feedback templates experience

Feedback templates in Studio now have card and list views with type-based filters (NPS, CSAT, Like/Dislike, Comment), a read-only summary page with a full-width interactive preview, and a single-page creation flow. Selecting a template opens its summary, and you can edit it from the overflow menu.

Arch AI is usable on mobile devices

The Arch AI interface now adapts to phone and tablet screens. On smaller displays, chat and artifacts fill the width behind a toggle instead of splitting the screen, and the status bar scrolls horizontally so the primary action stays visible.

Unsupported requests excluded from multi-intent cap

Requests that no agent can serve are no longer counted toward the multi-intent maximum, so they do not push valid requests into overflow. The Runtime Config panel also groups the multi-intent settings under clearer labels.

Warnings for invalid ESCALATE trigger formats

The compiler now warns when an ESCALATE trigger uses a format it cannot read, such as condition: or keywords: items instead of the required WHEN: syntax. Previously, these triggers were silently ignored, so the agent compiled without any escalation triggers, even though the author believed keyword escalation was configured. The new warnings (W355-W357) name the problem and its line number.

Clearer ENABLED_WHEN availability rules

Multi-line ENABLED_WHEN conditions that use the documented block indicator syntax no longer trigger a spurious warning. Missing-data handling in availability conditions is also corrected: a value compared with == null or IS NOT SET counts as known, while other missing values disable the route. The coordination UI for delegates and handoffs now reflects these rules.

Knowledge base

Business rules for knowledge base search

You can now add business rules to a knowledge base search tool in your agent’s ABL definition to control what a search returns, such as restricting results to certain sources, boosting values, or hiding internal content. Rules declared in an agent apply only to that agent’s searches, so two agents that share one knowledge base can get different results. Invalid rules on a project tool fail that tool’s load. Invalid rules in an agent’s override are skipped, and the agent falls back to the project tool’s rules.
Learn more →

TOOLS:
  search_policies(query: string) -> {results: object[]}
    description: "Search the HR policy knowledge base"
    type: searchai
    index_id: "hr-policies-v2"
    search_instructions: "Prefer the most recent policy version when duplicates exist."
    business_rules: |
      [
        {
          "id": "hide_drafts",
          "when": { "kind": "always" },
          "then": {
            "hide": [{ "field": "status", "operator": "eq", "value": "draft" }]
          }
        },
        {
          "id": "boost_official_pricing",
          "when": { "kind": "intent", "anyOf": ["pricing_question"] },
          "then": {
            "boost": [{ "field": "source", "value": "official_pricing_guide", "weight": 2.5 }]
          }
        }
      ]

Field mapping for knowledge base connectors

When setting up a knowledge base source, you can now map source fields to index fields in the connector wizard. This lets you control which document metadata is indexed and how it is stored, so search results reflect your content structure. Field mapping is available for the SharePoint, Confluence, Jira, Google Drive, Salesforce, ServiceNow, S3, SFTP, and Box connectors.

Per-object field scoping for Salesforce connector

The Salesforce Standard connector now lets you choose which fields to index for each Salesforce object. In the connector wizard, select objects, then refine the field scope per object so only the fields your agents need are indexed and searched. This reduces noise in search results and keeps your index focused on relevant content.

Enhanced field intelligence for knowledge bases

The field intelligence experience in knowledge bases now shows complete field capability metadata and offers an improved JSON field selection dialog, making it easier to map structured data fields to search indexes.

Voice experience

Grok realtime voice models

Grok realtime voice models from xAI are now available as a speech-to-speech voice provider. You can select Grok realtime when setting up a realtime LLM voice channel, alongside the existing OpenAI, Gemini Live, and Nova Sonic options.
Learn more →

Microsoft Speech deployment modes and Custom Neural Voice

Microsoft Speech services now support several deployment modes: Azure public cloud, Government, China sovereign cloud, Private endpoint, and Container. You can choose whether a credential is used for STT, TTS, or both, and configure Custom Neural Voice with a deployment ID and declared voices. The new settings appear in Voice Services and in each voice channel’s configuration.

Per-field speech recognition settings

Voice flow GATHER fields can now carry their own speech recognition configuration, so each field can use a different ASR provider, language, or vocabulary. For example, a name field can use a broad vocabulary, while an account-number field uses a tighter one. The deployment preflight validates these settings.

Extended voice filler sequences

Voice fillers now keep callers informed while Runtime processes responses, including during long-running operations and agent handoffs.

Voice fillers can now play up to eight lines per turn in the low-latency voice profile, up from five. On pipeline voice channels, fillers continue across agent handoffs and automatically replenish with up to two new lines when tool or model calls take longer than expected. Project-level settings take precedence over profile defaults.
For Gemini Live calls that use runtime-owned speech with the generative native filler, Runtime can play up to four short fillers per caller turn while preparing the response. Other Gemini calls continue to play one filler.

ON_START runs on OpenAI and Azure realtime channels

OpenAI and Azure speech-to-speech channels now run your ON_START handler and entry flow before the first turn, matching the behavior of Gemini Live channels. Previously, these providers skipped authored startup logic, so session initialization, language settings, and welcome messages set in ON_START did not take effect.

Queued caller speech is joined into one turn

When the agent is busy processing a response on KoreVG pipeline voice calls, additional caller speech is now combined into a single follow-up turn, instead of being handled as separate turns. Previously, the agent could respond separately to fragments such as “Hello?” and “Are you still there?” before addressing the caller’s actual question. Keypad input, active gathers, secure input, hold phrases, and language changes still keep their own turns.

Bind secure input results to session variables

The SECURE_INPUT CALL statement now accepts AS, RESULT_FIELDS, and SUCCESS_WHEN parameters. You can bind the delivery result to a session variable, choose which response fields are allowed out of the secure boundary, and define an exact-match success condition. Misplaced lines inside the block now raise a warning instead of failing silently.

Voice speech delivery is now a project setting

Streaming speech delivery on pipeline voice channels is now controlled by a Voice Speech Delivery toggle in project Runtime Config, replacing the previous environment variable. New projects have it on by default. Streaming delivery gives each response its own playback lifecycle, so the gateway resolves speech cleanly after barge-in, hold music, and filler transitions.
Learn more →

Inbound SIP header trust and transfer reliability

Voice channels now include an Inbound SIP Header Trust toggle that controls whether inbound SIP headers from the carrier are treated as SBC-asserted or untrusted. SIP REFER transfers with User-to-User URI headers are now correctly bracketed, resolving 488 rejections from some SBCs. The transfer trace also records whether the destination came from the agent definition or from a caller-supplied SIP header.

Post-transfer surveys from flow steps

Post-transfer surveys can now be configured in a FLOW step’s FEEDBACK block, without requiring an agent-level survey configuration. When a human agent closes the conversation, the platform uses the step-level survey if no agent-level survey is configured and either only one step defines a survey or all steps use the same survey template. Survey ratings now retain the survey template, agent name, and agent version, letting them appear correctly in per-survey Feedback Insights. Survey cards also show the survey expiry and the configured star or emoji rating style. The transfer record remains pending until the survey expires, then closes.

Welcome message barge-in control for voice channels

Pipeline voice channels now have an Allow caller to interrupt the welcome toggle in the channel configuration. When disabled, the agent’s welcome greeting plays fully before the channel accepts caller speech, so callers cannot talk over it. The setting applies only to channel-level welcomes and is preserved when you export and import the project.
Learn more →

Voice survey speaks the caller’s language

The post-agent CSAT survey on voice channels now plays in the caller’s language instead of the channel default. For example, a French-speaking caller hears the survey questions and closing lines in French, with a matching TTS voice, based on the agent’s locale catalog and language conditions.

Speech provider errors shown in session traces

Voice channels now accept custom speech settings. When a speech provider fails (for example, because of an invalid ASR configuration), the error appears as a clear, translated message in the session trace. Previously, these failures weren’t surfaced, making misconfigured speech pipelines hard to diagnose.

Clearer Gemini realtime voice settings

Studio now labels which filler and voice settings apply to Gemini realtime calls and which do not. The project’s Fillers-enabled and Voice-fillers switches are identified as master controls, and a notice appears on the Gemini channel panel when they are off. Proxy MOS is described as a ‘timing estimate’ on all surfaces.

Digital experience

Agent calls and screen sharing in the Web SDK

When a conversation is transferred to a human agent, the Web SDK now supports audio/video calls and screen sharing through the SmartAssist contact center. The SDK loads the WebRTC stack on demand and shows a shared media panel for these capabilities. Media events are relayed across runtime pods, so the feature works in multi-pod deployments.

Microsoft Agent 365 channel

Agents can now be deployed to Microsoft Agent 365 through a dedicated channel in the Deployments section. Configure the connection in its own setup tab, and the platform handles message routing between your agent and the Agent 365 ecosystem. The channel is available when enabled for your deployment.

Redesigned channel setup with inline guides

The Deployments section is now redesigned. Each channel connection shows a single setup guide with step-by-step instructions alongside the configuration form, replacing the previous tab layout.

Slack sender identity and per-thread sessions

The Slack channel now makes the sender’s user ID available to agent tools through session metadata, letting a tool look up the person who sent the message. Channel messages also get their own session per thread, so separate requests no longer share state. Direct messages continue as a single conversation.

LLM and models

New models: Claude 5, GPT-6, GPT-5.6, and Gemini 3.5 to 3.8

The model catalog now includes Claude 5 (Fable, Opus, Sonnet), GPT-6 (Sol, Luna, Astra), GPT-5.6 variants, Gemini 3.5 through 3.7 Flash, and Gemini 3.8 Flash as a text model. GPT-5 models are also available on AWS Bedrock, with a new Bedrock API key credential option. New Azure GPT-5.6 Luna connections default to the Responses API, which tool-calling agents need.

Project-level model parameter overrides

You can now override behavior parameters for each selected model at the project level. Set reasoning effort, max tokens, temperature, compaction threshold, and other parameters per model without affecting the shared workspace configuration. Only parameters the model supports are shown, and the runtime rejects overrides it cannot honor. The controls are on the Models page in project settings.

Model connections validated on creation

When you add a model connection, the platform now validates that the provider endpoint is reachable and the credentials work before saving. If validation fails, the platform cleans up the credentials and does not create the connection. Connections that can’t reach the provider are no longer saved in a broken state.

Connection labels for same-model credentials

When a model has more than one connection, each connection now shows a stable Connection 1 or Connection 2 label. The label appears in the Model Catalog, project, feature config, and eval model pickers, on admin connection cards, in Feature Config’s Resolves to column, and in the Why this model? section. Numbers never change or get reused when a connection is deleted. Models with one connection look the same as before.

Insights

Accurate turn timing and latency profile in Sessions

The Sessions trace view now shows accurate turn-level timings, including a latency profile, time to first token, and step-by-step timing offsets. Nested execution spans carry identity information, letting you trace exactly where time is spent. Lifecycle markers and session handoff events no longer inflate step durations.

Hook tool calls visible in traces

Lifecycle hook CALL actions now appear as separate tool call chips in the Observatory, showing the HTTP URL, request payload, response, duration, and error status. Previously, hook tool calls were hidden inside the hook row with no detail. The turn waterfall also groups tool calls correctly, letting you see the full sequence of model-initiated and hook-initiated calls in one view.

TTS first-audio latency in voice analytics

Voice interactions now capture the time from a text-to-speech request to the first audio byte reaching the caller. You can see this TTS first-audio latency in the Observatory interaction waterfall, alongside STT and LLM latency, to pinpoint where voice response delays originate.

Governance

JWT App credentials for API authentication

You can now authenticate to platform APIs using JWT App assertions, a new credential type alongside API keys and SDK tokens. Create a JWT App in Studio, configure key rotation, and optionally enable payload encryption. The platform verifies each assertion against your JWKS endpoint and protects against replay attacks.
Learn more →

Monitor and trace PII redaction enforced

PII fields configured for redaction are now consistently scrubbed from the Monitor live stream, workflow execution traces, and observability logs. Previously, certain paths didn’t apply the configured redaction rules, so sensitive values could appear in real-time monitoring views.
Learn more →

SSO just-in-time provisioning with role mapping

SSO configurations now support just-in-time (JIT) provisioning with automatic role assignment. When a user signs in through your identity provider for the first time, the platform creates their workspace membership and assigns a role based on rules you define from SSO claims. Existing members keep their current role, and the JIT policy always includes the owner domain.
Learn more →

Centralized voice gateway credentials

Platform administrators can now provision and manage voice gateway credentials centrally in the Admin console, then assign them to tenants. Projects use the platform credential by default, with an option to bring your own. Credential sets support rotation, tenant-level assignment, and retirement tracking, so you no longer need to configure voice credentials per project.

Secret config variables stay out of session state

Project configuration variables marked as secret are no longer copied into session state. They are held only in the runtime’s secure context and resolved at call time, so secret values never appear in execution snapshots, the Studio state stream, tool result embeds, or model prompts. Existing agents that reference secret config through template expressions still work; only the storage path has changed.

Workflow secrets excluded from execution traces

Secret configuration values used in workflow executions are no longer visible in execution traces or the Monitor. The runtime now passes secret references instead of decrypted values, so workflow traces show only the key names. This closes a path where secret configuration could appear in execution snapshots.

Audit events now attribute the acting user

Studio audit log entries for channel connections, SDK channels, webhook subscriptions, SearchAI mutations, auth profiles, and evaluation events now consistently record the user who performed the action and their IP address. Previously, some of these paths fell back to a generic system actor.

Richer model and connection audit entries

Model and connection changes in the Audit Log now include the model name, provider model ID, credential name, and before/after values for properties such as tier, default status, and display name. Previously, entries showed only field names or raw IDs, which made it hard to see what changed.

APIs and SDK

Standardized public Tool API

The public Tool API now uses three canonical paths, a standard response envelope, and a complete OpenAPI specification. The older per-environment paths are removed from the specification, so newly generated clients no longer include them. Creating a JWT App now requires a platform key with the sensitive-operations permission, and workflow invoke and status routes now check project scope.

Larger agent context for SDK sessions

SDK sessions can now pass up to 64 KB of agent context at bootstrap, removing the previous limit of 32 keys with 512-character values in customAttributes. The context is stored encrypted and delivered to the agent before the welcome message, so ON_START workflows have the full profile immediately.

Filter analytics sessions by caller number

The public analytics Sessions and Conversation History APIs now accept a callerNumber filter. Pass a phone number with its international prefix to retrieve sessions from that caller. Malformed or bare-digit numbers return a 400 error, and projects that have disabled caller identity in their reporting configuration receive a 403 error.

v1.7.3

Agent builder experience

Configuring multi-intent request completion

Two new settings for managing multi-intent interactions have been added to Runtime Config:

When a request is finished: Determines whether a request ends when the agent replies or only when the agent confirms it is done, so follow-up questions no longer move the conversation to the next request too early
Capturing every request in one routing decision: The supervisor lists all of the customer’s requests at once, so a second request isn’t lost when the model routes only one.

Auto strategy no longer runs requests in parallel

The auto multi-intent strategy now handles independent and dependent requests one after another, and asks the customer to choose when the requests are ambiguous. To run independent requests at the same time, set the strategy to parallel.

Clearer tool confirmation questions

When a tool call needs approval, the customer now sees a plain question built from the tool’s consent action and readable parameter labels, such as the account and amount. Sensitive values are masked, tool names and internal instructions are never shown, and you can change the wording in the agent’s MESSAGES block.
Learn more →

Model context for tool-hook responses

A tool’s on_result or on_error do: action can now include a short MODEL_CONTEXT beside its RESPOND. The customer still gets the full response, while later model calls see only your summary, keeping large responses out of the model’s own conversation context.
Learn more →

TOOLS:
  list_products() -> object
    MODEL_CONTEXT:
      FIELDS: [products[].id, products[].name, total]
    on_result:
      do:
        - RESPOND: "{{result.widget}}"
          MODEL_CONTEXT: "Displayed {{result.total}} products in order: {{result.product_summary}}."
          COMPLETE:
    on_error:
      do:
        - RESPOND: "Product search is unavailable."
          MODEL_CONTEXT: "Product search failed; no products were displayed."
          COMPLETE:

Project imports keep model selections

Project exports now include the project’s selected models. When you import the project or install it as a template, those models are linked, and the archive’s default model is applied. The import preview lists which models will be linked, and imported search pipelines keep their custom embedding model.

Knowledge base

Web crawler works with more protected sites

When a site answers the crawler with a browser challenge page, the crawler now retries with a browser profile. Sitemap crawls also keep your robots.txt setting, and recrawls keep their original settings.

Voice experience

Pre-written fillers continue long waits

On voice, the configured static filler pool now continues a long wait after the AI-generated lines run out, up to the per-turn limit, so callers no longer hear silence after the first filler. Custom filler prompts can produce up to three lines, the same as the built-in prompt, and a phrase is never repeated in one turn. Runtime Config shows an instruction you can copy into a custom prompt to get more than one line.

Recover unanswered Gemini caller turns

Gemini realtime channels have a new Recover unanswered caller turns only switch under Silent-turn recovery. When Gemini hears the caller but doesn’t answer within eight seconds, the platform restarts the session with the caller’s words. It doesn’t re-prompt after tool results.
Learn more →

Insights

Wider table widgets in custom dashboards

Table and detail widgets in custom dashboards can now save up to 200 columns, up from 20, so wide SQL results no longer need to be split across widgets.

v1.7.2

Agent builder experience

Control what the model sees from tool results

You can now decide how much of each tool result the model keeps in the conversation: the full result, selected fields only, or a short summary built from a template. Flow steps and completions also accept a note describing what the customer received, for later reasoning. The customer-facing response doesn’t change, and large tool payloads no longer crowd out the rest of the conversation history.

Idempotency header for HTTP tools is now opt-in

HTTP tools no longer send an idempotency header on every call by default, which resolves errors from providers such as Salesforce that reject unsupported headers. If your provider supports idempotency, such as Stripe, Adyen, or PayPal, enter its header name in the new Idempotency Header field in the tool’s resilience settings. The platform still protects retried calls from running twice.

Filtered bulk delete in Agent Tables

Agent Tables can now delete many rows at once by filtering on any column, not only a unique key. Grant the new bulk delete operation on a table tool, run DELETE ... WHERE in the query console, or use the API. A configurable row limit of up to 10,000 rows guards against accidental large deletes.

Duplicate-run protection for workflow and MCP tools

Workflow and MCP tools are now treated as having side effects unless you mark them otherwise, so retried calls no longer run the same operation twice. As a result, confirmation policies that apply to side-effecting tools now cover these tools too.

Clearer deployment preflight errors

Deployment checks now follow the same model rules as the runtime, so agents that work in testing no longer fail deployment with unexpected errors. When there is a real problem, the message names the agent and points you to the setting to fix, instead of showing a generic error count.

Oracle Database connector

Agents and workflows can now connect to Oracle Database 12.1 and later. Supported actions are Run Custom SQL, Insert Row, Insert Rows, Update Row, Delete Row, and Find Row, with results returned 500 rows at a time. Set up the connection with an auth profile, then use it from tools and workflows.

Knowledge base

FAQ connector

Knowledge bases can now ingest FAQ content through a dedicated FAQ connector. Upload question-and-answer pairs as a JSON or CSV file, or add them manually in a guided wizard that detects fields for you. You can exclude fields from indexing and split list values into separate entries. FAQ answers are also tuned so exact question matches rank well.

Voice experience

Caller transcription settings for speech-to-speech channels

OpenAI and Gemini speech-to-speech channels now include a Caller Transcription section where you can choose the transcriber, set the language callers speak, and list words to expect, such as brand names. Setting the caller language stops the model from guessing it on every turn, which could otherwise switch English speech into another script.
Learn more →

Continuous filler updates during long waits

Voice agents now prepare several filler lines at once: an opener, then timed progress updates spaced by a configurable follow-up gap after each line finishes. Callers hear steady updates during long lookups, instead of one filler followed by silence. Fillers follow the conversation’s language, and you can set a default or override language in Runtime Configuration.

Voice safeguards are now opt-in

The voice grounding guards, built-in guardrail bindings, and realtime channel defaults introduced in v1.7.0 are now off by default, so existing agents keep their previous behavior. You can enable them per agent with new execution settings. Realtime channels also gain an Advanced voice behavior section, with options such as which caller wording the agent acts on.

Step-level no-input prompts

Voice flow steps can now have their own prompts for when the caller stays silent, instead of sharing one agent-wide sequence. For example, a greeting step and a payment step can each respond to silence in a way that fits the moment. Learn more →

FLOW:
  ask_account:
    REASONING: false
    RESPOND: "What's your account number?"
    GATHER:
      - account_number: required
    VOICE:
      ON_NO_INPUT:
        timeout_ms: 10000
        messages:
          - text: "Could you tell me your account number, {{memo_name}}?"
    THEN: complete

Localized secure input prompts and tool arguments

Secure input prompts and tool call arguments can now use entries from your locale catalog, following the caller’s language with the usual fallback order. If a translation is missing, the flow receives a clear result it can handle, instead of raw template text.

More reliable tool calls on Gemini Live

Gemini Live agents no longer stall or lose tool results when the provider restarts its session mid-turn. If Gemini cancels a tool call and asks for it again after the caller speaks, the platform reuses the call already running, avoiding duplicate actions such as a repeated booking.

Governance

Payload encryption for public analytics APIs

Platform keys used with the Sessions, Conversations, LLM Logs, and Traces APIs can now encrypt request and response payloads. Turn on request encryption to get a key for encrypting what you send, and add your RSA public key to receive encrypted responses in a standard JWE envelope. The new Payload encryption section is in the platform key settings, and existing keys continue to work unchanged.

v1.7.1

Agent builder experience

Agent Tables on non-production deployments

Agent Tables can now be used from deployments in every environment, not only production. All deployment environments currently read and write the same production rows; per-environment table data is not yet available.

Voice experience

Background audio behavior with barge-in disabled

Background audio settings on Pipeline Voice channels now indicate when the audio actually plays. Between-sentence background audio requires barge-in to be enabled on the channel and a streaming text-to-speech voice. When barge-in is disabled, background audio plays only before the agent starts responding, because the reply is spoken as one continuous block with no gaps between sentences.
Learn more →

Caller identity on outbound calls

On outbound calls, the customer is now identified by the number the platform dialed, so contact, memory, and conversation history are attributed to the right person.

Emotech hotwords and Deepgram custom vocabulary controls

Custom vocabulary controls on voice channels now adapt to the speech provider you select:

Emotech channels add a Hotword Weight field that applies to every hotword in the list, with a recommended range of 5–10. Clearing the list turns hotwords off.
Deepgram channels show the vocabulary limit inline (up to 100 words) and warn when a saved phrase contains a comma, since the voice gateway splits comma-separated text into separate words.
Learn more →

Agent-spoken messages in the conversation transcript

Messages the agent speaks outside a normal turn — handoff announcements, no-input prompts, and farewell messages — now appear in both the transcript and the Conversation view. Previously, these reached the caller but were missing from the transcript, making voice sessions harder to review and audit.

Tool response scheduling default for Gemini Live

The Tool response scheduling setting on Google Gemini Live realtime channels now defaults to When idle, so the model finishes its current response before processing a tool’s final result. Progress updates for a pending tool call and native filler audio now always wait until Gemini is idle — regardless of this setting — so they no longer interrupt a response that’s already underway. Channels that relied on the previous immediate-response behavior can switch to Interrupt to restore it.

Governance

Runtime secret context for workflow tools

Workflow tools can now resolve secret references from the runtime secret context, as HTTP tools already do. The tool retrieves the credential at call time, and the secret value never enters the model prompt, execution logs, or traces — so it’s never exposed to the agent or persisted anywhere.

v1.7.0

Agent builder experience

Higher scenario and persona limits for evaluation suites

Evaluation suites can now include up to 500 scenarios and 10 personas, up from 3 and 3, letting you cover far more conversation paths in a single suite. The evaluator limit stays at 12 per suite, and suites generated automatically with Arch still start at 3 scenarios and 3 personas by default.

Unified human task handling

Process workflows can now handle approvals and data collection through a single Human Task node. It pauses the workflow, notifies assignees, and can collect structured data through a reusable Form before resuming based on the selected outcome. You can define SLAs, configure reminders and escalation steps, and specify how to handle SLA breaches. Forms can be created once as reusable, project-level assets and used across workflows. This gives process workflows a simpler way to manage human decisions, data collection, and task deadlines while reducing duplicate form creation.
Learn more →

Scenario-level session variables in Evals

Evaluation scenarios can now define their own session variables, not just personas. When a scenario and a persona share a variable, the scenario’s value wins by default, unless the persona marks that key as priority, in which case the persona’s value is kept. This lets you test the same persona across different contexts without cloning it for each one.
Learn more →

Conditional routing in ESCALATE blocks

ESCALATE routing can now be an ordered list of destinations, each guarded by a condition such as channel or language and evaluated in order, with a required default as the last entry. Previously, routing to different destinations per channel — for example, a voice transfer versus a digital handoff — required duplicating the entire agent. Now, one agent can express both.

Runtime-Guard for realtime orchestration

A new Runtime-Guard mode lets a realtime model stay in control of the conversation, proposing actions and extracting values naturally, while the runtime validates every context write and business action against the active flow step, and rejects anything invalid, stale, duplicate, or out of order. This keeps realtime conversations flexible without giving up control over side effects; the stricter runtime-owned mode remains available where it’s still needed.

Knowledge base

Knowledge base included in import/export

Project export now includes each knowledge base’s configurations, including source configurations and pipeline stages, instead of just an empty tool shell pointing at a missing index. On import, the knowledge base, its search index, and its sources are provisioned automatically, and the knowledge base tool is reconnected to the new index.

Voice experience

PCI-compliant secure DTMF input with voice channel suppression

Flow authors can now collect sensitive DTMF input, such as credit card numbers, while fully suppressing the voice channel: no speech reaches ASR models and no audio is captured in call recordings during collection.

Amazon Nova Sonic (S2S) support via Bedrock

Amazon Nova Sonic is now available as a realtime speech-to-speech voice provider via AWS Bedrock, alongside the existing OpenAI and Gemini Live options, with barge-in and live transcription support.

Configurable static filler messages per language

Static filler messages played during voice processing pauses can now be added or updated per language directly from Studio’s Runtime Configuration panel, with no platform deployment required.

Call-status reason fields for unanswered outbound calls

Agents and campaign workflows can now branch on why an outbound call ended - busy, no answer, network timeout — instead of only knowing that it failed.

Low-latency customer-service voice profile

Voice projects can now apply a low-latency profile tuned for customer-service calls, using faster model tiers and disabled reasoning by default, in exchange for quicker responses. Studio flags when a project’s settings deviate from the recommended values and lets you apply them individually or all at once.
Learn more →

Voice channel recovery on LLM and tool failures

The platform now classifies voice and tool failures into specific recovery codes, such as an LLM timeout, an unavailable tool, or a provider disconnect, instead of collapsing everything into a generic failure, with configurable retry and fallback behavior per channel. When a tool call is running slowly, the runtime can also generate a response speculatively, so the caller hears something sooner instead of dead air.

Insights

Credential attribution in LLM logs

LLM Logs now record which credential or auth profile served each model call, so you can attribute token and cost consumption to a specific source rather than just a project or model. This includes distinguishing two credentials that share a display name across projects. Learn more →

Governance

Improved AWS KMS BYOK support

AWS KMS bring-your-own-key configurations now support AssumeRole authentication, matching the structured setup Azure KMS already had, and no longer silently fall back to the platform’s own AWS credentials if a tenant’s are incomplete.
Learn more →

Auto-accept workspace invites option for SSO

Workspace admins can now enable automatic invite acceptance for SSO sign-ins, so a user invited to a workspace no longer hits an authentication error if they log in through their identity provider before separately accepting the invite. The option is configurable per workspace under SSO Settings.
Learn more →

GCP Workload Identity Federation support

GCP Workload Identity Federation is now supported for LLM and voice credentials, letting projects with a no-static-keys security policy authenticate to Gemini and Vertex AI, including Gemini Live for voice, without a long-lived service-account key file.

v1.6.5

Knowledge base

PDF formula extraction

Knowledge bases can now extract mathematical formulas from PDF documents. Previously, formulas weren’t recognized as either text or images during extraction, so they were left out of the indexed content. They’re now captured along with the rest of the document. This is enabled by default for newly created knowledge bases, improving retrieval quality for technical and scientific content. Existing knowledge bases keep their current extraction configuration.

SharePoint connector: quick actions and permission-aware search

The SharePoint connector’s overview screen in Studio now includes a Quick Actions bar for Sync, Pause, Resume, Edit Config, Re-Auth, Health Check, and Search, all directly from the connector overview.
Learn more →

Voice experience

Gemini Live: configurable context-window compression

For long voice conversations on Gemini Live, you can now configure context-window compression directly from the channel’s configuration panel in Studio — turn compression on or off, set the token count at which it triggers, and the token count it compresses down to. This helps longer conversations retain relevant context without hitting the model’s context limit.

v1.6.4

Voice experience

Contact Centre lifecycle events from CCAI

The platform now derives Contact Centre lifecycle events (agent connected, agent disconnected, ACW completed) from CCAI’s conversation_info payload, removing the need for explicitly pushed event entries. Canonical transfer_initiated events are now emitted for voice SIP transfers. The dispositionCode and wrapUpNotes field mapping has been corrected, and close events now include queue and assignment fields.

Language-based voice channel routing

Callers can now select their preferred language at the start of a call (via DTMF, for example, press 1 for English, 2 for Spanish) and are routed to a language-specific voice channel with matching ASR/TTS settings. The selected language is retained for the entire session, preventing the agent from unexpectedly switching between languages mid-conversation.

Runtime-driven dynamic language switching

Callers can now switch languages mid-conversation with a simple spoken request (for example, “continue in French” or “speak Greek”), instead of only selecting a language once at the start of the call. Runtime validates and owns the language state, so switches only happen on clear, explicit requests — casual mentions like “I have a French account” won’t trigger a change. The selected language is preserved consistently across handoffs, reconnects, and session resumption, and voice fillers (for example, “one moment”) now play back in the caller’s active language instead of defaulting to English. This also extends language-consistency support to Gemini-based realtime voice sessions, alongside the existing pipeline voice channels.

Dynamic Welcome Messages for Agent Assist

The Kore Agent Assist channel now renders templated ON_START welcome messages using live session data. By initializing the session and running ON_START setup routines (SET, CALL) upfront, the platform ensures authored responses populate with real values. Static welcome message configurations for agents remain fully supported and unaffected.

Transcript grouping timeout

Flow authors can now configure a transcript grouping timeout, controlling how long the platform waits before grouping transcript segments into a single turn.
Learn more →

Gemini 3.8 Live voice model support

Gemini 3.8 Live is now available as a realtime voice model option.
Learn more →

Governance

Deployment lifecycle audits

Studio now surfaces deployment lifecycle audits, giving visibility into the deployment process for governance and compliance. Deployment audit metadata now includes additional context, and catalog compatibility is preserved across audit entries.
Learn more →

v1.6.3

Agent builder experience

Comment-only survey templates and simplified follow-up rules

Feedback templates in Studio can now be authored without a numeric rating, so you can collect comments on their own. Per-rating follow-up questions can also inherit their visibility rule from the survey’s primary control, instead of being configured separately.

Voice experience

Per-turn TTS delivery policy for protected voice responses

Flow authors can now configure the STREAMING delivery mode for voice responses, alongside the existing barge_in control. This lets authors protect important spoken responses — such as compliance disclosures or handoff completion messages — from being interrupted, truncated, or affected by tool calls, reconnects, or mid-call capability changes.
Learn more →

v1.6.2

Voice experience

Amazon Nova Sonic speech-to-speech voice

Voice channels can now use Amazon Nova Sonic, a speech-to-speech (S2S) model served through AWS Bedrock, as a realtime voice provider. Configure it in Studio’s Voice Services page, using either access-key or IAM-role-based authentication.
Learn more →

Vertex AI Gemini Live for realtime voice

Realtime LLM voice channels can now be powered by Vertex AI’s Gemini Live models, in addition to the direct Google Gemini Live provider. Model resolution and credential handling for Vertex use the same Voice Services configuration flow as other realtime providers.
Learn more →

Native multilingual voice profiles

Voice channels now support per-language voice profiles: each supported interaction language can have its own ASR and TTS provider and voice selection, with the voice runtime handling automatic cutover and fallback. Learn more →

Voice CSAT survey on the return leg

Voice transfers over SIP now support a post-agent CSAT survey played back on the call’s return leg after a human agent completes the interaction, using a reserved x-callee-number SIP header to identify the transfer target.

GCP authentication for Vertex AI and Gemini

Auth profiles for Vertex AI and Gemini now support:

Google Cloud Workload Identity Federation (WIF): Authenticate without long-lived service-account keys.
GCP service-account authentication: Configure Vertex AI auth profiles with an uploaded GCP service-account key and test the connection.

Gemini Live JSON tool results and non-blocking tool fillers

Gemini Live channels now return tool results in JSON mode and can play non-blocking filler audio while a tool call is still in progress, reducing dead air during longer tool executions.

Warning on imported model selection

Studio now warns when a project selects an imported model that may not be fully compatible with the current configuration, so teams can catch mismatches before deployment.

Digital experience

Post-transfer CSAT survey

After an agent-to-agent transfer, the Web SDK can now present an end-of-transfer CSAT survey to the user, with responses captured through a new webhook survey resolver.

v1.6.1

Governance

Realtime model visibility and health monitoring

Workspace administrators now have clearer visibility into which models are available for realtime voice interactions across projects. The platform also continuously verifies realtime model connectivity through an internal health-monitoring service, helping ensure configured realtime models are available for voice use.
Learn more →

v1.6.0

Agent builder experience

Per-server rate limits for MCP tools

Rate limits can now be configured per MCP server, with Studio controls to set the limit and automatic enforcement across tenants, so one server’s traffic can’t overwhelm the rest. Learn more →

Automatic indentation validation and repair in the ABL editor

The ABL editor now detects indentation problems as you write and offers an auto-fix, with clearer diagnostics when something can’t be repaired automatically.

Link eval suites to a version, deployment, or environment

Eval suites can now run against a specific project version, deployment, or environment. This helps teams determine whether a score change came from a project edit or another change, and evaluate what is actually deployed.
Learn more →

Workflow versions, settings, and a redesigned monitor experience

Workflows now support versioning and a dedicated settings tab, alongside a redesigned monitor experience in Studio with clearer run detail, timeline, status, outcomes, and payload views.
Learn more →

Move workflows between projects

Workflows can now be exported from one project and imported into another, with a preview of any conflicts before the transfer completes.
Learn more →

Customize webhook trigger request and response formats

You can now customize the structure of webhook trigger requests and responses. Use the default format or define a custom format for the request payload the workflow receives and the response payload it returns.
Learn more →

See why an agent isn’t ready to run

Studio now surfaces structured readiness diagnostics before a chat starts, explaining what’s blocking execution — such as a validation issue or incomplete topology — instead of only failing at runtime.

Higher-fidelity execution token budget alerts

Studio’s debug panel now warns when a run’s execution approaches its token budget, with the limit enforced consistently from the compiler through the core parser — so you catch runaway prompts during testing instead of in production.

Redesigned architecture review experience

The Arch AI architecture review now shows staged review progress as it works, clearer proposal summaries, and suggested next steps once you create a review — making it easier to follow what Arch AI checked and what to do next.

Import and export projects using a public API

Projects can now be imported and exported using a public API authenticated with a platform key and processed asynchronously. This helps external systems and pipelines move projects across deployments without using the Studio.
Learn more →

Improved model resolution

Model resolution now supports non-reasoning models for named-entity recognition tasks, with additional provider options and an improved resolver for query-time model selection.

Agent tables: durable structured data store

Agent tables are a new structured data store you define in Studio, which agents read and write at runtime through a built-in table tool. Built on Postgres, tables are a durable, typed, queryable store for state that needs to persist and be shared across channels and sessions, rather than living only inside a single conversation.

Knowledge base

Scheduled web crawls for SearchAI

SearchAI can now crawl configured websites on a recurring schedule, not just on demand, with a Studio panel to set up and manage crawl schedules. Learn more →

Exact SharePoint site targeting

The SharePoint connector for SearchAI lets you scope to specific sites via SharePoint’s ‘Sites.Selected’ permissions, featuring a Studio interface for direct site selection rather than requiring broader site-collection access. Learn more →

Voice experience

Richer handoff context for ServiceNow, Five9, and Salesforce agent transfers

Agent transfers now resolve handoff metadata at runtime and project it into SIP headers for ServiceNow, with matching adapter support for Five9 and Salesforce, so more conversation context reaches the receiving system. Learn more →

SIP termination URI support for BYOC voice provider

Bring-your-own-carrier voice channels can now be configured with a SIP termination URI, including provisioning and cleanup handled from Studio.

Author realtime voice delivery policy directly in Studio

Studio now lets you author the delivery policy for realtime voice sessions — controlling how responses are delivered as the conversation happens — directly in the agent editor, instead of only through the underlying ABL definition.

Reference SIP data in voice transfer destinations

Voice call transfer destinations can now include {{sip.}} placeholders that resolve to SIP data at runtime, with validation to make sure only the permitted values are forwarded.
Learn more →

Scoped barge-in control for voice interruptions

Barge-in behavior (whether a caller can interrupt the agent mid-response) can now be scoped to an individual step or flow instead of only a project-wide setting, giving you finer control over which parts of a voice agent are interruptible.

Outbound dialing from voice agents

Voice agents can now place outbound calls, not just receive them, using a new public API that triggers a call on demand over telephone or SIP through a configured voice channel. Each call is tracked with an ID and logged for later review.

Support for phone numbers with a leading plus sign

BYOC voice channels now accept E.164 phone numbers with a leading plus sign by default, closing a gap where such numbers could be rejected.
Learn more →

Choose an OpenAI speech-to-text model for voice channels

Voice channel configuration now lets you select which OpenAI speech-to-text model to use, in addition to the existing provider options.
Learn more →

Bring your own custom speech provider

Tenants can now register and use their own custom speech provider for voice channels, alongside the platform’s built-in speech provider options.
Learn more →

On-premises Deepgram support for voice channels

Voice channels can now connect to an on-premises Deepgram deployment using shared gateway speech credentials, instead of only Deepgram’s cloud endpoint.
Learn more →

Connect greeting audio now plays by default

You can now upload and manage the greeting audio callers hear when a voice channel connects, with a preview before publishing. New deployments now play a connection greeting by default, rather than requiring you to enable it.

Filler noise for more natural-sounding voice agents

Voice agents can now play configurable filler noise, such as ambient background sound, with its own playback settings, so pauses feel less abrupt.

Speech-to-Speech (S2S/V2V) model support

Speech-to-speech (S2S) voice models now go through the same Model Hub selection path as every other model, instead of being locked to workspace-level Voice provider settings. This enables per-project keys and configuration for S2S, matching text and speech-to-text models.

Digital experience

Verified user identity for Slack channel connections

Slack channel connections can now enforce a policy requiring verified Slack user identity before a conversation proceeds, configurable from Studio.
Learn more →

Improved async handoff for AI for Work channels

AI for Work (AI4W) channel connections now support overriding activity visibility for streamed events, with an improved resumption service for conversations that hand off asynchronously.

Consent support for omnichannel live sessions

The embedded web SDK widget now supports collecting user consent during live, omnichannel sessions, improving the handoff experience across channels.
Learn more →

New channel: Exchange Graph Direct email

Email is now supported as a deployment channel through Microsoft Exchange Graph, with mailbox management, message erasure support for compliance, and a dedicated channel editor in Studio.

Upload attachments without presigned URLs

File attachments can now be uploaded as base64 content and delivered via a signed URL, giving developers an alternative to handling presigned URLs directly.

Insights

Full feedback survey templates

You can now build feedback surveys from templates with configurable scoring, condition rules, and localization, and view the results through a dedicated analytics dashboard. Agents can also collect surveys by voice, and a public API exposes them.
Learn more →

PII controls for custom analytics dimensions

PII policy can now govern custom analytics dimensions. The platform enforces the policy wherever it writes or displays that data, and a new Govern PII card in Studio shows policy status.
Learn more →

Filter sessions by end-user phone number

Use the caller phone number filter to trace a specific caller’s history across sessions, rather than searching session by session.
Learn more →

Custom date ranges for Sessions and Traces

The Sessions and Traces pages now support a custom date range picker, matching the one already available on the Insights page.

New session and KPI metrics

Analytics now includes a unique-sessions metric with support for ungrouped aggregates, and workspace-level KPI weighting and error-rate calculations are now more accurate.

Governance

Workspace-level controls for model access

Workspace administrators can now audit and enforce which AI models a tenant can use, with idempotent controls to manage model access consistently across a workspace.
Learn more →

Smoother SSO invite and lockout recovery

Pending workspace invites are now accepted automatically when a user signs in through SSO, instead of requiring a separate acceptance step, and memberships locked due to a lockout are automatically restored once the lockout expires.
Learn more →

OAuth2 Resource Owner Password Credentials auth profile

Auth profiles now support the OAuth2 Resource Owner Password Credentials grant type, available consistently across the runtime, compiler, and Studio, for integrations that need this flow instead of an authorization-code or client-credentials grant.

External secrets via Azure Key Vault

You can now connect an external Azure Key Vault as a secret source, with runtime routing to resolve secrets from it and a Studio UI to connect and browse available secrets, instead of storing every credential directly in the platform
Learn more →

Guard against accidental encryption provider switches

A new safeguard prevents accidentally switching between platform-managed and bring-your-own-key encryption for a workspace’s KMS configuration.

DTMF input can now be classified as PII

Keypad (DTMF) input during a voice session can now be explicitly classified as PII, with policy enforcement applied the same way it is for other PII types.
Learn more →

Organization-wide proxy egress

Organizations can now configure a proxy for all outbound traffic across the runtime, Studio, compiler, and editor, with built-in connectivity testing, policy invalidation when settings change, rate limiting, and observability into proxy behavior.
Learn more →

v1.5.4

This release includes bug fixes and platform improvements.

v1.5.3

This release includes bug fixes and platform improvements.

v1.5.2

Agent builder experience

Support for Base64-encoded file attachments

Alongside the existing multipart/form-data upload path, the platform now accepts Base64-encoded attachments in application/json requests, using filename, mimeType, and contentBase64. Successful uploads can return a short-lived signed delivery URL. If the delivery URL expires, you can request a new one without re-uploading the file.

Workflow Agent node accepts attachment IDs

The Agent node’s canvas attachments field now accepts attachment IDs directly, in addition to https URLs, so a workflow can hand an already-uploaded attachment to an agent node without re-hosting it as a URL first.
Learn more →

v1.5.1

This release includes bug fixes and platform improvements.

v1.5.0

Agent builder experience

Artemis CLI

A new customer-facing CLI is now available, letting developers automate the full Agentic lifecycle, including authentication, project sync, eval runs, and deployment, from external systems, without Studio.
Learn more →

Microsoft Foundry model catalog support

Microsoft Foundry is now available as a unified provider, replacing the previous Anthropic-only integration. Select models from multiple publishers, including OpenAI, Meta, Mistral, Cohere, and Anthropic, or enter custom deployments. The configuration also supports Azure AD Service Principal authentication with automatic token discovery.

Recurring evaluation run schedules

Eval runs can now be scheduled to recur daily, weekly, or monthly, with time zone selection and flexible start and end conditions.
Learn more →

Exact timestamps on eval runs

Eval runs in Studio now display exact start and completion timestamps alongside the existing relative time labels, across Studio.

Unsaved-changes warning in the ABL editor

Studio now warns you before you lose unsaved work when leaving the agent editor, with the option to save, discard, or stay.

In-app Help Center and first-time user welcome experience

Studio now includes a built-in Help Center with guided documentation, a platform primer, feature highlights, and support resources. It also offers a unified welcome screen for first-time users with quick-start actions—creating a project, using a template, connecting Artemis MCP, inviting team members, and accessing docs and support—reopenable anytime from the profile menu.

Knowledge base

Improved OCR and language handling for SearchAI document extraction

SearchAI now automatically detects a document’s language and selects the corresponding RapidOCR model, rather than relying on a single fixed model regardless of script. This improves results for Arabic, Thai, Urdu, and Latin-script scans.
Mixed-language documents are kept in their source language rather than translated.
SearchAI detects pages with unusable or sparse text (scanned images, glyph-name artifacts, collapsed text, or vertical right-to-left output) and runs a full-page OCR fallback to recover readable text.
Explicit language and OCR-language settings now carry through the SDK, connectors, and every extraction path (sync, async, split-PDF, streaming URL, extraction-only), so configured language hints are honored consistently.
Automatic OCR now runs only when a PDF’s embedded text layer is genuinely unusable, is disabled for pure-text formats by default, and caps fallback retries rather than retrying indefinitely.
A visual enrichment or OCR failure no longer causes the whole document to fail — extraction continues downstream with any usable text.

Connector enhancements: Custom base URLs and OAuth authentication

Multiple connectors now support custom base URLs and expanded OAuth 2.0 Client Credentials authentication, enabling connections to on-premises deployments and sovereign cloud tenants instead of relying only on public cloud endpoints:

Jira: Custom base URL and product-version selection for Jira Data Center deployments
Confluence: Custom base URL, OAuth 2.0 Client Credentials authentication, and product-version selection
SharePoint: Custom base URL and OAuth 2.0 Client Credentials authentication through custom Azure AD endpoints, including sovereign clouds such as .de
ServiceNow: Custom base URL, OAuth 2.0 Client Credentials authentication, and product-version selection
Salesforce: OAuth 2.0 Client Credentials authentication for non-interactive, scriptable authentication
Microsoft 365 (Excel): OAuth 2.0 Client Credentials authentication for non-interactive Microsoft 365 authentication

Voice experience

Voice Playground for softphone-based voice testing in Studio

Playground now supports voice testing alongside text testing, allowing developers and QA teams to initiate a softphone interaction, speak with the agent, view the live transcript and call state, and open the session in Observatory for detailed voice debugging. Learn more →

Voice call transfer and escalation to a live agent

The Escalation block in Studio now supports voice call transfers, in addition to the existing chat hand-off to an agent desk. Voice calls can be escalated to a human agent by integrating the call_transfer tool with the ESCALATE block; previously, escalation supported only chat-based handoff. Developers can choose a call transfer type (SIP, PSTN, or relay) and a transfer method (invite, refer, or bye/release-link) directly in the Escalation editor, to prevent failure messages from being announced twice. Learn more →Voice agents can also transfer calls to external phone numbers, not just SIP endpoints, by configuring a PSTN phone number as an agent transfer destination in the project voice settings. Learn more →

Genesys Cloud Audio Connector for voice bots

Voice calls can now connect to Genesys Cloud through the Audio Connector, using the AudioHook protocol, with support for both Pipeline and Realtime voice modes.
Learn more →

Configure language-specific TTS profiles for multilingual voice agents

Voice Pipeline channels can now define an explicit TTS profile for each configured language, including provider, voice, model, output format, pace, and temperature. When the interaction language changes, the agent switches to the configured profile for that language, instead of inferring a voice or provider from the default configuration. Learn more →

Amazon Connect live-agent transfer via SIP BYE with generated conversation summary

Voice calls routed from Amazon Connect to an Artemis agent can now be transferred to a live agent, along with a generated conversation summary. The platform then releases the bot’s call leg via SIP BYE, allowing Amazon Connect to resume its contact flow and route the caller with context intact.
Learn more →

Validation of STT timeout configuration against provider-specific limits

Voice channel configuration now validates the speech/segmentation-silence timeout against the selected STT provider’s actual supported range (for example, Microsoft’s 100–5000 ms), instead of accepting any positive value, and rejects invalid values rather than silently clamping or forwarding them.

Forward custom outbound SIP headers on call transfer

Projects can now configure a set of allowed outbound SIP headers to forward during voice call transfers. Administrators select which predefined headers (for example, x-queue, x-skill) a project permits, and agents can attach values for those headers when initiating a SIP REFER or INVITE transfer. Enforcement, sanitization, and drop telemetry are applied at runtime. Learn more →

Multi-gateway BYOC SIP support

BYOC (Bring Your Own Carrier) voice channels now support configuring multiple SIP gateway IP addresses and CIDR ranges, replacing the previous single-IP field. Each entry is provisioned as its own Jambonz SIP gateway, with full lifecycle management for create, update, rollback, and cleanup.

Digital experience

Genesys Cloud Open Messaging integration for digital channel agent handoff

Conversations on Web Chat, WhatsApp, and other messaging channels can now be handed off to a live agent through Genesys Cloud.
Learn more →

Add Support for Microsoft Teams and Copilot channel

Microsoft Teams and Copilot are now supported as a new deployment channel. Users can create a Teams channel instance from the Studio deployments screen, configure credentials and identity policies, and have the runtime enforce a verified Teams identity on inbound messages.
Learn more →

File attachment support in the API channel and web SDK chat widget

The file attachment option in the chat widget can now be shown or hidden per project, and the API channel now passes along caller-supplied context to the agent.
Learn more →

Auto-disconnect idle chat widget sessions

The embedded chat widget can now be configured to automatically disconnect or end a session after a period of user inactivity.
Learn more →

Insights

One place to see every LLM call

A new Generations view shows every model call, from Studio, search, and live traffic, in one place, with cost, latency, and token usage.
Learn more →

Filter analytics by environment and channel

Analytics dashboards is now filetered by environment (development, staging, production) and channel, so testing data doesn’t get mixed in with production numbers.
Learn more →

See who gave feedback

The feedback table and exported reports now show which user submitted each piece of feedback, plus any custom tags.
Learn more →

Public analytics APIs for sessions, traces, and LLM usage

External systems can now pull conversation history, session and session-summary reporting data, messages, execution traces, and LLM cost and usage ledger data directly through public APIs.

Governance

Simpler guardrail threshold configuration and scanner configuration

Guardrail policies now provide a simpler way to configure detection thresholds at the project level. Learn more →
Workspace administrators can configure guardrail providers and their available scanners, while project authors can select which configured providers the project can use. Guardrail providers also support a wider range of classifier response formats. Learn more →

Entity-based PII detection and anonymization

PII detection and anonymization now support entity-based detection. Authors can create PII patterns using entity recognition (built-in or model-backed) instead of regex. An ignore-values list lets authors exclude known values from detection and reduce false positives.
Learn more →

Auth Profile fields to reference deployment environment variables

Project Auth Profile fields — Token URL, Refresh Token URL, Scopes, Client ID, and Client Secret — can now be populated from environment variables defined under Deployments, instead of requiring these values to be entered directly into the Auth Profile.

OAuth and PKCE support for remote MCP servers

Remote MCP server authentication profiles now support OAuth configuration, including authorization, token, and refresh URLs, scopes, and PKCE. PKCE can be enabled with the S256 challenge method.

Simpler Azure encryption setup

Customers using their own Azure encryption keys no longer need to register an app in their Azure tenant — the platform now manages that identity, and key rotation is simpler. This also adds support for additional compliance certifications (GDPR, SOC2, NIST, FIPS-140-3).
Learn more →

v1.4.4

Voice

Author SIP transfer destinations directly in the agent DSL

You can now declare a trusted SIP transfer destination directly on the built-in call_transfer tool, instead of relying solely on the project-level transfer target. The destination is authored as a non-model default (default_source: config), so it’s never visible to or fillable by the LLM. The compiler validates it at publish time — catching blank values, unresolved templates, and malformed SIP URIs before an agent goes live.
Learn more →

Configure a minimum STT confidence threshold

You can now set a project-level minimum speech-to-text (STT) confidence threshold for voice interactions. Transcripts below the threshold are treated as no input, preventing low-confidence speech recognition results from triggering agent responses. Set the threshold between 0 and 1; set it to 0 to disable it.
Learn more →

Tools

Build dynamic batch payloads with HTTP body templates

HTTP tools now support {{#each input.X}} in body templates, letting you dynamically generate batch payloads from array inputs in a single request. You can also use session secrets in the generated payload, while keeping secret values protected from the model.
Learn more →

Security and Compliance

Simplified Azure BYOK setup with a platform-owned identity

Azure Bring-Your-Own-Key (BYOK) encryption for same-tenant Azure Key Vault no longer requires customers to register or maintain their own Entra app registration or a federated credential configuration. Artemis now uses a platform-owned Workload Identity Federation application — customers grant it a scoped Key Vault role and provide their key location, with no client ID, secret, or certificate ever entered or stored.
Learn more →

New KMS compliance controls for BYOK

Workspaces using BYOK/KMS can now select a compliance level (GDPR, SOC2, or NIST), which the platform enforces at the point of any KMS configuration write. Key rotation is now split into two distinct, independently controlled actions — DEK retirement and KEK rewrap — each with its own preflight check before running.
Learn more →

v1.4.3

This release includes various updates to enhance the stability and resilience of agent handoffs, long-duration calls, and authentication flows, as well as improvements to conversation routing and recording playback.

v1.4.2

This update includes new features and enhancements summarized below.

Tools

Secure secret handling for tools and workflows

Tools can now return sensitive values — API keys, tokens, credentials — without exposing them to the LLM. These values are captured into an encrypted, session-scoped vault and kept invisible to the model and the conversation transcript. Use the new {{_secret.X}} syntax to restore a captured secret into a downstream tool’s HTTP headers or request body.
Learn more →

Voice

Configure a PSTN transfer target number

You can now configure a dedicated Public Switched Telephone Network (PSTN) transfer target number for agent-initiated call transfers, alongside the existing SIP transfer configuration. Outbound PSTN transfers now correctly present the provisioned caller ID.
Learn more →

Channels

New API scope for closing sessions

Integrators using the API channel can now issue a platform API key scoped specifically to closing sessions (close-session), instead of requiring a full administrative key to terminate a conversation programmatically.
Learn more →

Microsoft Teams: verified identity and thought streaming

Microsoft Teams is now supported as a verified-identity channel, with identity enrichment, auth-profile-based credential mapping, and a new channel setup experience in Studio. Teams conversations can also now stream the agent’s thinking or progress updates in real time, with a new streaming toggle in the Teams channel configuration.
Learn more →

Knowledge base

Box and Amazon S3 connectors

You can now connect Box and Amazon S3 as knowledge sources, making it easier to index and search content stored in these repositories. These connectors enable seamless ingestion of documents from your cloud storage into your knowledge base.

v1.4.1

This update includes new features and enhancements summarized below.

Workflows

Set a default value for data entry fields in Workflows

Workflow builders can now set the default value of fields in a Data Entry node using either static text or expressions. Expressions can reference data from workflow triggers or previous workflow steps. For example, a reviewer’s form can be pre-filled with an AI-generated draft for approval or editing.
Learn more →

SDK

Control attachment visibility in the Web SDK

The Web SDK Theme Builder now has a toggle to show or hide the attachment icon in the chat widget’s compose bar. This lets FAQ and informational AI agents offer a cleaner, text-only interface when file uploads aren’t needed, while keeping attachments enabled by default elsewhere.

Analytics

Control SQL query operations for public analytics widgets

Admins can now allow or restrict specific query operations for public-facing analytics dashboard widgets, from a new SQL Validation section in the Public Widget Data Policy settings. Supported restrictions include SQL comments, multiple SELECT statements, JOIN clauses, common table expressions (CTEs), set operations (UNION, INTERSECT, EXCEPT), and OR conditions.
Learn more →

v1.4.0

Feature enhancements and new capabilities introduced in this release are summarized below, followed by the breaking changes that may affect existing projects — review the migration steps before recompiling or republishing them on 1.4.0.

Agent building (ABL/Studio)

Rich-content buttons can open a link on click

Message ACTIONS buttons and carousel card buttons can be configured with a URL so that clicking it opens the link directly in a new tab, instead of posting an action back to the agent. It is supported on the Web SDK/widget, Telegram, and LINE; other channels automatically fall back to a normal button.
Learn more →

Git integration for agent projects

Developers can connect a Git repository to a project to sync agent definitions and localization assets with version control, including push and pull from the Studio editor and conflict-resolution tools for simultaneous edits.
Learn more →

Inline ABL editor on the agent canvas

You can now edit agent logic directly in an inline code editor without leaving the visual canvas.

Voice

Configurable caller-silence (“no-input”) handling for pipeline voice agents

Voice agents on the KoreVG/Jambonz pipeline can now set re-prompt messages, retry counts, and timeout windows for silent callers, with locale-aware prompts and a final action (hang up, custom message, or transfer).

Custom vocabulary for speech-to-text

You can tune speech recognition to recognize domain-specific terms such as product names and jargon. It leads to improved accuracy in transcriptions.

Filler-agent context and language routing for IVR

Filler (“thinking”) audio during voice calls now carries conversation context and follows the caller’s active language.

No more silent gaps during backend tool calls

Updates to filler and no-input handling with Google Gemini Live now reduce dead air, repeated speech, and lost language context while a voice agent waits for a backend tool call. The updates require no configuration change.

Channels

Microsoft Teams: user-identity and channel-template support

Microsoft Teams connections can now resolve caller identity and add user details like email address, display name, or user ID to the context.
Learn more →

HTTP Async (webhook) channel: interruption and streaming support

The HTTP Async channel now supports mid-response interruption and streamed delivery of agent responses.
Learn more →

Multiple message rendering

Agents can now send several distinct messages in one turn, each rendered as a separate bubble instead of one combined block.

APIs and SDK

Agent transfer uses webhook channel now. Agent transfer supports transfering attached documents. The allowed file formats and file size limits are defined in project-level attachment settings.SDK sessions now receive a runtime-controlled activityVisibilityPolicy that decides updates of which customer-facing activity to show. It lets you control fillers like Searching..., safe thought summaries, action progress, system notices, and redacted tool previews. Raw reasoning, secrets, and unsafe tool data is hidden.

Analytics and Insights

Public API for saved dashboard-widget data

External tools, BI connectors, and scheduled exports can now fetch the data behind any saved custom-dashboard widget through a public REST endpoint, without re-implementing its underlying query. It supports both synchronous requests and asynchronous jobs with polling and cursor-based pagination.

PII entity preview

You can now preview which entities in a conversation will be detected and masked as PII before publishing an agent.

Integrations and Single Sign-on

Workspace sign-up, SSO, and access requests have new controls

Tenantless users are now routed to onboarding instead of being left in an ambiguous state, workspace SSO is separated from org SSO, and sign-up or workspace creation can be gated by feature flag. Review these settings after upgrading, especially if your org uses SSO or gated sign-up. Invited and password-recovery SSO users are now routed through separate flows, and Azure SSO app selection and Azure BYOK metadata are represented as structured fields instead of opaque JSON — review these if your workspace uses Azure SSO or BYOK.

Workflows

Workflow enhancements

Workflows now have better support for array-typed inputs/outputs, per-workflow rate limits, scheduled (cron) triggers, a JSON step editor, and an improved workflow overview page.
Learn more →

LLM and Model Configuration

Adaptive thinking config in Anthropic models

The platform now supports adaptive thinking as a configuration for Opus 4.7 and Opus 4.8 models.

Breaking changes and what to check before moving to 1.4.0

The changes below affect existing projects when you move to 1.4.0. Review the items relevant to your role developer or admin, and resolve them before promoting or republishing projects on the new version.

For agent developers

Supervisor role is now explicitcanRespondDirectly is deprecated as the primary way to set supervisor behavior. New agents should declare BEHAVIOR: role: router (route only) or role: orchestrator (combine results and respond). Existing agents keep working through compatibility mapping, but a role that conflicts with the old boolean setting now fails to compile.

Before (1.3.3):
    SUPERVISOR: Benefits_Router
    canRespondDirectly: false
    HANDOFF:
      - TO: Claims_Agent
    	WHEN: intent.category == "claims"
After (1.4.0):
    SUPERVISOR: Benefits_Router
    BEHAVIOR:
      role: router
    HANDOFF:
      - TO: Claims_Agent
    	WHEN: intent.category == "claims"

Routers can no longer own returned workA router can only select the next agent. If the parent needs to use a child agent’s returned result, declare it an orchestrator instead. Agents that use DELEGATE inside a router will no longer compile.

Before (1.3.3) — DELEGATE inside a router, now a compile error:
    SUPERVISOR: Benefits_Router
    canRespondDirectly: false
    DELEGATE:
      - TO: Plan_Lookup_Agent
    	WHEN: intent.category == "benefits"
    	RETURNS:
      	plan_summary: object
After (1.4.0) — declared as an orchestrator instead:
    SUPERVISOR: Benefits_Coordinator
    BEHAVIOR:
      role: orchestrator
    DELEGATE:
      - TO: Plan_Lookup_Agent
    	WHEN: intent.category == "benefits"
    	RETURNS:
      	plan_summary: object
    INSTRUCTIONS: |
      Use plan_summary to answer the member. Do not expose unrelated session memory.

Compiler warnings can now block publishingDiagnostics that used to be silently accepted — conflicting role and flag settings, routers containing DELEGATE, unmapped handoff inputs, unsupported tool array schemas — now stop a project from compiling or publishing. Review and resolve any outstanding warnings before promoting existing projects on 1.4.0.Handoff and delegate inputs must resolve, or the call failsA handoff or delegate no longer runs with a missing or stale input value. Add an explicit MEMORY read for any value the child needs, and use ON_RETURN to map back the results the parent depends on.For deterministic FLOW parents, HANDOFF with RETURN: true now suspends at the transition boundary and resumes from the saved next step once the child returns — review verifier, auth, eligibility, and multi-agent flows where a parent depends on a child’s returned values.

Before (1.3.3):
    HANDOFF:
      - TO: Authentication_Agent
    	WHEN: intent.category == "auth"
    	RETURN: true
    	INPUT:
      	memberId: '{{member_id}}'
After (1.4.0):
    MEMORY:
      member_id:
    	type: string
    	access: read
    HANDOFF:
      - TO: Authentication_Agent
    	WHEN: intent.category == "auth"
    	RETURN: true
    	INPUT:
      	memberId: '{{member_id}}'
    	ON_RETURN:
      	action: resume_intent
      	map:
            authentication_status: result.authentication_status
        	member_id: result.member_id

Tools must declare the memory they read and writeA tool that needs conversation memory must declare context_access.read and/or context_access.write. If the declared memory isn’t available, the tool doesn’t run — check tools that previously assumed memory was always present.

Before (1.3.3):
    TOOLS:
      get_plan_information(memberId: string) -> { success: boolean, plans: array }
    	type: http
    	endpoint: "/eligibilitymembersinfo/query"
    	method: POST
After (1.4.0):
    MEMORY:
      authentication_status:
    	type: string
    	access: read
      member_plan_summary:
    	type: object
    	access: write
    TOOLS:
      get_plan_information(memberId: string) -> { success: boolean, plans: array }
    	type: http
    	endpoint: "/eligibilitymembersinfo/query"
    	method: POST
    	context_access:
      	read: [authentication_status]
      	write: [member_plan_summary]

Declare the item type for array inputsWorkflow array inputs still work without an item type, but declaring one — for example itemType: "string" — is now recommended. Workflow and connector array parameters are also projected into the model-facing tool schema with an items definition in 1.4.0, so some model providers (for example Gemini/Vertex) can reject the whole tool declaration if a typed array doesn’t declare its item type.

Before (permissive array):
    { "name": "available_days", "type": "array", "required": true }
After (typed array):
    {
      "name": "available_days",
      "type": "array",
      "itemType": "string",
      "required": true
    }

Review PII detection settingsNew projects start with no built-in PII detection enabled by default (enabled_builtins: []). If your agent relied on default PII protection, explicitly enable the detection types you need before publishing.Re-check SearchAI metadata and title filtersFilters that were previously dropped or ignored — unknown standard filters, metadata filters, and title-only filters — are now honored correctly. Review saved knowledge base and tool filters, since results may return different, more accurate documents after upgrading.Gather validation outputs are now canonicalFLOW gather validation and retry-exhaustion fields are now lowered into the canonical execution model instead of relying on display-only alias names or implicit retry fallbacks. Make gather validation outputs explicit and don’t rely on a label as if it were a runtime variable name.Terminal output is held until the final response is confirmedRuntime now defers final output until response checkpoints pass, which cuts down on duplicate terminal responses and output arriving after cancellation — but it also surfaces incorrect output-ownership wiring earlier. Re-test multi-agent flows with async fan-out, remote handoff returns, delayed voice responses, or structured COMPLETE payloads without text.Review agents that branch on channel or transfer stateMessage metadata, channel identity, and transfer state now persist more consistently across reconnects, cancellation, escalation, and remote handoff returns. If an agent branches on channel identity or resumes after a remote or human-agent transfer, re-test it on 1.4.0.

For admins

Recompile and re-import migrated projectsProjects moved between tenants or environments now go through stricter namespace validation before they can be published. Re-import or recompile any migrated projects on 1.4.0 and confirm there are no outstanding namespace warnings.Validate workflows triggered outside of StudioStart-input defaults and required-input validation for workflows are now applied server-side across every trigger path — cron, webhook, connector-polling, and agent-triggered — not just Studio’s Run dialog. Validate promoted workflow versions on each of these trigger paths, not only by running them manually in Studio.

v1.3.1

Voice

Improved Microsoft streaming TTS reliability

Voice channels using Microsoft streaming text-to-speech now provide more reliable voice playback following periods of silence, reducing interruptions and helping ensure agent responses continue to stream as expected.

New voice provider: Sarvam AI

Sarvam AI is now available as a native speech-to-text and text-to-speech provider with tenant-scoped credential management.
Learn more →

v1.3.0

Multi-agent and Arch AI

Real-time Arch AI activity feedback

The Arch AI activity panel now shows live, readable updates on each action as it happens, instead of delayed or raw output.
Learn more →

Arch loop evaluation and optimization

Arch AI now supports an iterative eval-optimize loop, allowing agents to be evaluated and refined through repeated automated passes, including voice-agent evaluation scenarios.
Learn more →

Arch AI engine tracing

Arch AI now produces structured trace events for each internal step of agent generation, optimization, and evaluation, giving visibility into Arch’s multi-step reasoning process for debugging and optimization.
Learn more →

Voice diagnostics knowledge for Arch AI

Arch AI can now offer diagnostic guidance for voice issues like missing audio, TTS failures, or STT misconfiguration.

Persistent memory: Append, unique, and max-size modifiers

Agent memory (REMEMBER STORE) can now append values to arrays, with optional deduplication and a size cap, rather than overwriting them. This makes it easier to build agents that accumulate history, like preferences or past interactions, over time.
Learn More →

Voice

Call recording

Voice agents now support call recording through the Kore Voice Gateway, with project-level recording configuration and enterprise recording policy enforcement. Recordings can be stored in Azure Blob Storage in addition to the platform’s default storage.
Learn more →

Outbound DTMF tones

Voice agents can now send outbound DTMF tones during a call, enabling interaction with IVR systems and other DTMF-driven telephony endpoints.

Genesys Cloud integrations

Artemis now integrates with Genesys Cloud on two fronts: the Audio Connector for live voice calls routed from Genesys Cloud, and Open Messaging for digital/text conversations routed through Genesys Cloud’s contact center.
Learn more →

Voice transcript sync and pre-transfer history on handoff

When a voice agent hands off a call to Kore SmartAssist/CCAI, the conversation transcript is synced to the receiving system, and a new API exposes the full pre-transfer conversation history, giving the receiving agent complete context.
Learn more →

Voice agent transfer via Kore Voice Gateway

Voice agents can now transfer live calls via SIP through the Kore.ai Voice Gateway, carrying conversation context to the transfer target.
Learn more →

Custom vocabulary and keyword boosting for speech-to-text

Voice channels now support custom vocabulary and keyword boosting for speech-to-text providers, improving recognition accuracy for domain-specific terms — configurable at the project level and per deployment channel.

Locale-aware agent responses

Reasoning agents can now reference project locale assets directly in their ABL responses, enabling multi-language agent output driven by the session locale without hardcoding translations.
Learn more →

Channel-appropriate response formatting

A single reasoning agent can now reliably produce channel-appropriate output—Markdown formatting for web and natural spoken text for voice—from one agent definition, without needing separate agents per channel.

Workflows and integrations

SOAP/WSDL integration support

Workflows and tools now support SOAP integrations: upload a WSDL file, browse its available operations, and register them as SOAP-type HTTP tools for agents and workflows — with a dedicated WSDL asset repository that supports versioning.

Integration actions as agent tools

Pre-built connector actions (for example, Jira or Slack operations) can now be added directly as agent tools — pick a connector and action from a catalog, and the platform handles authentication and parameter mapping automatically.

Webhook transport for contact-center handoff

Artemis now supports a webhook-based transport to hand off conversations to CCAI/SmartAssist and exchange ongoing customer messages during a live escalation.
Learn more →

Knowledge base and Search AI

New packaged connectors: Confluence, ServiceNow, SFTP, and Google Drive

The knowledge base connector catalog now includes packaged connectors for Confluence Cloud (with permission-aware delta sync), ServiceNow, SFTP, and Google Drive (with folder-level selection), in addition to existing custom HTTP connectors.
Learn more →

Custom script pipeline stage

SearchAI ingestion pipelines now support a custom ‘Transform Script’ stage, letting authors write sandboxed JavaScript to transform document content and metadata during ingestion.

Search query debugging controls

The knowledge base Search Test page now includes independent toggles to disable vocabulary resolution and query preprocessing, making it easier to isolate and debug individual stages of the query pipeline.
Learn more →

Knowledge base export in reusable agent modules

Reusable agent modules can now include knowledge base content (indexes, connector configuration, and source metadata) when published, so consumer projects reference the source module’s indexed content without duplicating storage.
Learn more →

Security, identity, and access

Workspace-aware enterprise SSO

Enterprise SSO now supports workspace-level configuration, allowing each workspace to maintain its own SAML/OIDC authentication settings independently of the organization-level SSO configuration, with new admin governance controls.
Learn more →

Custom roles

Workspace owners and admins can now create custom roles with granular permissions beyond the built-in role set and assign them to project members for fine-grained access control.
Learn more →

Expanded Azure Key Vault BYOK authentication options

Bring Your Own Key (BYOK) with Azure Key Vault now supports three authentication methods: typed credential fields (replacing raw JSON configuration), certificate-based service-principal authentication with a pre-save connection test, and Workload Identity Federation for secretless, cross-tenant access.
Learn more →

Per-tenant Platform-Managed key governance

Platform administrators can now view and manage Platform-Managed (Kore-owned) encryption key configuration on a per-tenant basis, with server-side enforcement and audit logging of key lifecycle events.

Custom PII redaction with capture-group backreferences

Custom PII redaction patterns now support regex capture-group backreferences (\1 / $1), enabling partial redaction that preserves non-sensitive parts of matched text — for example, masking all but the last four digits of an account number.
Learn more →

Analytics and insights

Output-item analytics

A new analytics subsystem captures structured data items extracted from agents’ final responses and includes a Studio authoring UI for defining output-item analytics configurations.

Enhanced failure alerting

Aggregate live-session failure alerts now include failure-type summaries.

Custom metrics and dimensions for workflow analytics

Workflow analytics now support custom metrics and dimensions, allowing teams to define and query workflow-specific measurements throughout the analytics pipeline.
Learn more →

Dashboard data export

Analytics dashboard widgets can now be exported as CSV directly from the Studio UI for offline analysis and reporting.

Export feedback

Collected user feedback (ratings, comments, and session references) can now be exported as a .csv file from the Insights → Feedback page for offline analysis.

Per-widget dimension filters

Custom Dashboard widgets now support per-widget dimension filters, allowing different widgets on the same dashboard to apply independent custom-dimension filters.
Learn more →

Administration

Vertex AI via GCP service-account authentication

The platform now supports Google Cloud Vertex AI models — both managed (MaaS) and dedicated-endpoint deployments — using GCP service-account credentials with automatic token refresh, configurable directly in Studio.
Learn more →

Auth profiles can reference environment variables

Auth profile fields can now reference project environment variables (for example, {{env.CLIENT_SECRET}}) instead of hardcoding credentials, so auth profile definitions can be shared safely across environments.

Usability improvement

Active workspace indicator

The account menu now displays the name of the currently active workspace, making it clear which workspace you’re operating in when switching between multiple workspaces.

1.2.2

Public-key JWE authentication for Hosted Exchange Web SDK

Hosted Exchange Web SDK now supports public-key JWE authentication, enabling secure authentication using customer-managed public keys while maintaining compatibility with existing authentication flows.
Learn more →

Environment-specific SAML Entity IDs

Each deployment environment now uses a unique SAML Service Provider (SP) Entity ID, allowing administrators to configure separate SAML applications within their identity provider (IdP) for development, staging, and production. Existing SAML SSO configurations must be updated with the new SP Entity ID after upgrading to this release.

v1.2.1

Arch AI enhancements

Arch AI now provides a more reliable experience for authoring, generation, and evaluation. Improvements include support for larger source documents, more reliable preservation of handoffs across multiple sources, refined validation to reduce unnecessary generation failures, and non-blocking diagnostics to help identify issues without interrupting the generation process.

Language-aware ON_START messages

The Web SDK session interaction language is now accessible within the ON_START event, enabling language-specific welcome messages without requiring workflow-based workarounds. This simplifies multilingual agent design, reduces unnecessary latency, and allows startup logic to use standard ABL conditional expressions.

v1.2.0

Multi-agent orchestration

Enhanced delegation and handoff workflows

Multi-agent workflows now support both returning and non-returning delegation patterns, enabling more flexible interruption, escalation, and transfer scenarios. DELEGATE and HANDOFF behaviors are now distinguished, making it easier to implement workflows where conversations either return to the originating flow or are permanently handed off to specialized agents.

Smarter Supervisor direct responses

Supervisors can now respond directly to user requests by default, reducing unnecessary routing between agents. This ensures delegated responses are returned to users while still allowing Supervisors to operate as routing-only orchestrators when needed via the canRespondDirectly setting.

Web SDK

Web SDK multi-language UI customization

The Web SDK chat widget now supports language-based customization of UI elements such as chatbot names and typing hints, making it easier to support multilingual deployments.

Simplified Web SDK integration

Web SDK deployments can now be configured with public API keys and origin allowlists, eliminating the need to generate per-channel embed tokens and simplifying setup with greater control over embedded deployments.

Web SDK distribution via NPM and CDN

The Web SDK is now available through npm and CDN distribution channels, making it easier to integrate into modern web development workflows and deployment pipelines. The SDK is available here.

Voice

Enhanced SIP header support for voice channels

SIP header information from inbound and transferred voice calls is now available, enabling routing and business logic based on telephony metadata — including queue information, customer identifiers, language preferences, campaign data, and transfer context.

Security and compliance

OCI Vault support for Customer-Managed encryption keys

Bring Your Own Key (BYOK) now supports Oracle Cloud Infrastructure (OCI) Vault Key Management Service (KMS), in addition to AWS KMS, Azure Key Vault, Google Cloud KMS, and external REST-compatible KMS, providing greater flexibility to manage customer-controlled encryption keys and compliance requirements.

v1.1.4

Search AI and knowledge management

Enhanced Search Tool configuration

Search Tool parameters, such as filters and retrieval settings (for example, topK), can now be configured as mandatory or optional, providing greater control over knowledge retrieval behavior and ensuring that required parameters are consistently applied during tool execution — improving retrieval accuracy and reducing reliance on prompt-based instructions.

Custom HTTP connector for knowledge bases

Content can now be ingested into a Knowledge Base from custom HTTP-based data sources. The connector supports scheduled synchronization, schema discovery, field mapping, and secure connectivity options including mTLS, making it easier to integrate proprietary enterprise systems and external data sources with RAG applications. Ingestion workflows also support JavaScript-based content transformations, allowing data to be preprocessed before chunking and indexing.

AI engineering tools

Expanded model support

The Platform now supports Anthropic’s Claude Opus 4.8 model.
Learn more →

Other improvements

SSO enhancements

The platform now auto-generates SAML and OIDC endpoint URLs for identity provider configuration. These system-managed URLs ensure consistent setup, reduce manual errors, and are available only to organization owners and admins.
Learn more →

v1.1.3

This release includes bug fixes and usability improvements across the platform.

v1.1.2

This release includes bug fixes and usability improvements across the platform.

v1.1.1

Updates to Web SDK experience

Improved the Web SDK experience with welcome-message lifecycle enhancements, theme refinements, markdown rendering improvements, and footer-link fixes. Enhanced the Web SDK session handling by preserving and securely resuming idle sessions instead of recreating conversations.

Updates to SearchAI and knowledge management

Simplified the SearchAI Knowledge tool DSL, improved knowledge-tool binding and resolution behavior, and enhanced structured-data processing and knowledge ingestion reliability.

Updates to tool and gather definitions

Added support for the effect and provides properties in AgentTool definitions and the satisfiedBy property in GatherField definitions.

v1.1.0

Credential management updates

Introduced Config Variables ({{config.X}}) and Auth Profiles as the new credential management model, replacing Tool Secrets and the Tool Secrets Management API.

New memory and handoff capabilities

Added support for the per_activation reset mode for session variables. Enhanced handoff workflows with ON_RETURN.set, ON_RETURN.map, and typed handoff history support.

Validation and tool execution enhancements

Added batched LLM validation for fields using VALIDATION_PROCESS: LLM and clarified last_<tool_name>_result storage behavior, including store_result semantics.

Project import and tooling improvements

Added support for tool-stub synthesis during project bundle import, allowing tool definitions to be generated from inline agent TOOLS: signatures and re-exported automatically.

ABL syntax and field updates

Updated the ABL expression syntax with method-call operators such as input.contains(), .matches(), and abl.is_number(). Updated the human-agent escalation fields to use user.wants_human_agent.

v1.0.4

Updates to filler messages

Enhanced Filler Messages with contextual filler history support. Administrators can now configure how much recent conversation history and how many previously generated filler messages are used when generating new filler responses. The filler prompt also supports the new {{conversationHistory}} and {{previousFillers}} placeholders for more context-aware filler generation.

v1.0.3

Tool readiness evidence in Arch AI

Arch AI now shows readiness evidence for the tools it adds while building your agents. Each tool is clearly labeled as backed by a live integration or by a stub/mock placeholder, and a project health view surfaces what still needs to be configured before the agent is ready to go live.

More reliable agent generation from SOPs

Generating agents from Standard Operating Procedures (SOPs) now produces more complete agent topologies. Arch AI avoids incomplete builds and validates the generated structure for quality, so SOP-derived projects start from a fuller, more accurate design.

Authoring evidence in Arch AI reviews

Arch AI authoring reviews now display evidence badges and incorporate evaluation results. Authoring guidance for constructs such as MEMORY, HANDOFF, TOOLS, GATHER, AGENT, SUPERVISOR, and guardrails is rendered from the Platform’s own construct definitions, giving you clearer, more consistent recommendations as you build.

A smoother Arch AI chat experience

Several reliability improvements make working in the Arch AI chat smoother: errors that have already been resolved (such as model or provider configuration issues) no longer linger in the interface, you can start a new chat even when a previous session ended in an error, and sending a routine message no longer triggers a full refresh of the interface.

DELETE method for HTTP tools

HTTP tools now support the DELETE method, alongside GET, POST, PUT, and PATCH. Your agents can now call REST endpoints that remove resources directly from an HTTP tool.

v1.0.2

Agents and Arch AI

Evidence-grounded authoring in Arch AI

Arch AI now grounds its authoring review in real parser and evaluation evidence, surfaces authoring-automation health and improvement opportunities, and routes proven best practices into generated agents — so the guidance you see while building reflects what the compiler and evals actually report.

Stronger evaluation gating

Invalid eval imports are now marked non-runnable, runs with zero coverage are blocked before they start, cancelled runs are preserved, and the winning evaluation status is reported clearly — making evaluation results more trustworthy.

More reliable agent generation

Improved the reliability of the Arch AI build experience: blueprint generation recovers from blank turns, BUILD topology failures are now recoverable instead of terminal, and the chat/generation experience handles error states more gracefully.

Canvas delegate edges render correctly

Fixed a regression on the agent canvas where delegate relationship edges were drawn as handoff edges. Delegation and handoff relationships are now visually distinct.

Behavior Profile honors all CONVERSATION fields

The Behavior Profile CONVERSATION block now applies all configured fields, resolving a case where only style and tone were applied and other fields were silently dropped.

Tools and workflows

Tools page “New Tool” menu fix

The Studio Tools page ”+ New Tool” continues to be visible when the current tab has no tools. Now, you can create any type of tool regardless of which tab is active.

ABL workflow triggering from the AI4W channel

Fixed an issue where ABL workflows didn’t trigger when invoked from the AI for Work (AI4W) channel.

Insights and dashboards

Agent Transfer dashboard

Added an Agent Transfer insights dashboard. You can now analyze how and when conversations are transferred between agents.

Voice

Reliable voice channel provisioning

Fixed an issue to improve the reliability of voice channel setup, during Jambonz WSS application provisioning.

Studio and administration

Consistent Audit Log date formatting

Audit Logs now use a consistent date format between the From–To filter and the results table, along with display fixes for smaller screen resolutions.

Git integration fixes

Git-sync failures are now surfaced clearly in Studio, and import-error messages are normalized for easier troubleshooting.

Security and privacy

SDK origin allowlist enforcement

The Web SDK public-key embed now respects the selected channel’s origin allowlist.

Hardened SearchAI downloads and outbound requests

Strengthened security across SearchAI file downloads and outbound webhooks (including redirect handling) and hardened rich-content sanitization in the Web SDK.

Credential scrubbing in debug traces

Tool credentials and other secrets are now scrubbed from debug traces and compile-time trace events, preventing sensitive values from appearing in diagnostics.

Cleaner euser error handling

A failed LLM response is no longer misclassified as an empty response, and internal developer diagnostic strings aren’t displayed to the users.

v1.0.1

Agents and Arch AI

Reusable agent modules

Package agent logic as modules and reuse them across multiple projects instead of rebuilding per project. Imported tools now route correctly into the agent’s TOOLS: section.Learn more about Catalog Settings → and Create and Manage Projects →.

Arch AI — one-shot generation and contract-driven authoring

Arch AI adds a one-shot generate endpoint (describe an agent and get a working blueprint), and its generated constructs are now derived from the compiler contracts with enforced typed topology and validated tool contracts — producing more reliable, deploy-ready agents.
Learn more →

Runtime guardrail enforcement

Guardrails configured on an agent are now reliably applied at run time, closing a gap where policies could be skipped during live conversations.
Learn more →

Knowledge and Search AI

Clearer web-crawl ingestion

The knowledge web-crawler setup now gates the wizard so you can only advance when the source is ready to crawl, and a crawl indexing lifecycle with a live status strip shows ingestion progress from submission to searchable content.

Knowledge fast-path search

A faster knowledge-retrieval path for agent chat, with the schema-preview endpoint now accepting JSON Schema for type extraction — lower-latency answers in conversation.
Learn more →

Custom pipeline authoring redesign (contract-first)

A redesigned authoring experience for custom indexing/processing pipelines with contract-first validation, making pipeline configuration clearer and less error-prone.
Learn more →

Crawler reliability improvements

Fixes to web-crawl ingestion (including docling filename/extension handling) improve the reliability of content ingested into knowledge.
Learn more →

Voice

Unified voice provider registry and capability matrix

A single registry of voice (STT/TTS) providers and their capabilities gives consistent provider selection and configuration across voice agents.
Learn more →

Configurable Gemini Live VAD sensitivity

Gemini Live voice-activity-detection (VAD) sensitivity is now exposed in the speech-to-speech channel configuration, letting teams tune turn-taking and interruption behavior.
Learn more →

Channels, contact center, and connectors

Contact-center “After Call Work” (ACW) handling

The runtime now receives and processes ACW/disposition events after a human-agent call closes (emitting acw_completed and agent_disconnected with SmartAssist event data, keeping sessions alive to capture disposition) — enabling accurate post-call workflows after an agent transfer.
Learn more →

AI-for-Work channel type

A dedicated channel lets AI for Work invoke ABL agents directly.
Learn more →

Service Management connector category

A new connector category groups Zendesk and ServiceNow under Service Management.
Learn more →

LLM and models

AWS Bedrock provider integration

Adds AWS Bedrock as an LLM provider with IAM role support and additional connection and configuration options.Learn more about AI Configuration → · Supported AI Models →

Studio, insights, and analytics

Insights Auditing — automated metric analysis and recommendations

Automated analysis of agent and quality metrics with improvement recommendations, wired across the agent-performance, quality-monitor, and customer-voice dashboards.
Learn more →

Studio feedback viewer

A recent-feedback table with filters in Studio to review captured user feedback.
Learn more →

Studio navigation redesign and design-system standardization

Refreshed Studio navigation and a standardized design system, including sidebar localization labels.

Security, administration, and runtime

Platform domain allowlist and DB-managed Platform admins

Adds a configurable Platform domain allowlist and database-managed Platform admin roles for tighter access governance.
Learn more →

JWE encryption for Hosted Exchange WebSDK auth

Adds JWE support to secure the Hosted Exchange WebSDK authentication flow, with enforced JWE policy scope and hardened browser-SDK CORS handling.
Learn more →

Response metadata: LLM-generated versus scripted

Responses now carry metadata indicating whether a reply was LLM-generated or scripted, surfaced over the SDK/WebSocket structured content.

Tools and WebSDK

URL-encoded request body support

HTTP tools now support application/x-www-form-urlencoded request bodies, with the body type selection preserved.
Learn more →

2xx responses return the response body

Non-2xx HTTP responses now return the response body instead of being treated as outright execution failures, so agents can act on error payloads.
Learn more →

Rich-text rendering

The WebSDK now renders hyperlinks, paragraph formatting, and lists instead of showing them as plain text.
Learn more →

v1.0

Kore.ai Agent Platform { Artemis } is now available, with the following capabilities:

Agent Blueprint Language (ABL™) — create and manage agents.
Arch™ AI — generate agents from SOPs.
SearchAI — manage and retrieve knowledge.
Channels — deploy agents on web and voice.
Tool integrations — connect agents to external tools.
Insights dashboards — monitor agent performance.

For details, see Platform Overview.