> ## Documentation Index > Fetch the complete documentation index at: https://koreai.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # Decision models Decision models use small, fast models to handle specific decisions during agent execution, such as selecting a specialist to handle a request or determining whether a user has made one request or multiple requests. By handling these checks before calling the full AI model, decision models can reduce latency and model usage. If a decision model is uncertain, the platform falls back to the full AI model. Decision models can be configured at the following levels: | Where | Who can configure | What you can do | | - | - | - | | Admin Console | Platform administrator (super admin) | Enable Decision models for a workspace. | | AI Configuration > LLM Providers > Model Catalog | Workspace administrator | Add and configure decision providers, configure authentication, and view provider usage. See [Configure decision models for a workspace](#configure-decision-models-for-a-workspace). | | Project > Settings > Runtime Config | Project administrator | Select providers for supported decision steps and configure tiers, thresholds, budgets, and other step-level settings. See [Configure decision models for a project](#configure-decision-models-for-a-project). | > [!NOTE] > > Decision models are disabled by default at the workspace, project, and individual-step levels. Enabling the feature at the platform level makes the configuration available but does not automatically enable it for every project or decision step. ## Configure decision models for a workspace Workspace administrators configure decision providers that projects can use to make fast, predefined decisions during agent execution. To add a decision provider: 1. In your workspace, go to **AI Configuration > LLM Providers > Model Catalog**. 2. In the **Decision providers** section, select **Add decision provider**. 3. Configure the provider settings. | Field | Description | | - | - | | Name | A unique name for the provider. Choose the name carefully because it cannot be changed later. | | Adapter | The protocol adapter used to connect to the provider. Select **OpenAI Decisions** for an OpenAI Decisions endpoint or **Typed judgment (System One)** for TypeSafe or another System One endpoint. | | Endpoint | The provider endpoint URL. Public endpoints must use HTTPS. Private endpoints may require additional approval. Confirm the approval requirements with your platform administrator. | | Model | The model identifier expected by the provider. This field is optional. | | Timeout (ms) | The maximum time allowed for a provider call, from 20 to 5,000 ms. The default is 200 ms. | | Model context (tokens) | The optional context-size setting. Recent conversation turns are trimmed, oldest first, to fit the request. The current message is always included. Leave the field empty to use the default limit of 4,000 characters. | | Authentication | Choose one of the available authentication options: **Use a model connection credential:** Reuse an API key from an active, workspace-level model connection. OpenAI Decisions accepts only an OpenAI model connection credential. **Use an auth profile:** Use a shared, workspace-level API key or bearer profile. Select the required profile from the list. Do not include API keys in the endpoint URL. | | Circuit breaker | The circuit breaker temporarily skips a provider after repeated failures. **Failure threshold:** The number of consecutive failures that trigger the circuit breaker. The default is 5. **Reset timeout (ms):** The time to wait before resetting the circuit breaker. The default is 30,000 ms. | 4. Select **I acknowledge data egress to this provider** to confirm that you understand the data sent to the configured endpoint, then select **Create provider**. The decision provider is now configured and available for selection in your project. ## Configure decision models for a project Project administrators can choose which decision steps use a decision provider and configure the settings for each step in their projects. For descriptions of the available settings, see [Runtime configuration](https://koreai.mintlify.site/agent-platform/administration/general-settings.md#reasoning-pipeline). **Prerequisites**: Before configuring decision models for a project, ensure that: * A platform administrator has enabled Decision models in the Admin Console. * A workspace administrator has configured at least one decision provider. To configure decision models in your project: 1. Open the project in Studio and go to **Settings > Runtime Config**. 2. Configure the required settings in the following sections. | Runtime configuration section | Setting | Description and steps | | - | - | - | | **Reasoning pipeline** | **Routing provider** | Configures a decision provider for routing decisions. Ensure that **Enable Pipeline** and the classifier pipeline are enabled to use this setting. Enable **Decide with a provider**, add a configured provider under **Tiers**, and configure the available settings as needed. | | **Conversation handling** | **Confirmations** | Resolves yes/no answers to confirmation prompts. Enable **Decide with a provider**, add a configured provider under **Tiers**, and configure the available settings as needed. | | **Conversation handling** | **Gather responses** | Resolves answers to gather and choice prompts. Enable **Decide with a provider**, add a configured provider under **Tiers**, and configure the available settings as needed. | | **Conversation handling** | **Flow escape** | Detects when a user wants to leave the current flow. Enable **Decide with a provider**, add a configured provider under **Tiers**, and configure the available settings as needed. | | **Conversation handling** | **Provider dispatch** | Configures how decision providers are called and the minimum confidence required when a decision step does not specify its own threshold. Configure the available settings as needed. | | **Conversation handling** | **Scores** | Configures weighted scores evaluated by fan-out providers. These scores can be used in conditions through the `decision_scores` object. For example, `decision_scores.escalation.value` accesses the escalation score. | The runtime tries providers in the configured order. If no provider returns a sufficiently confident answer, the LLM handles the decision. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.