## Summary Enterprise administrators need the ability to control which AI models can be used by Duo Agent Platform. Currently, an admin can pin a single model and then end-user model selection is disabled. End-user model selection is currently available in agentic chat. ## Problem Enterprises need model governance for: - **Cost control** - Allow models that fit their budgets - **Compliance** - Restrict to models/providers that meet data residency requirements - **Quality** - Ensure only high-capability models are used for critical workflows - **Consistency** - Standardize on specific models across organization ## Proposed Solution Add administrative controls for model configuration: 1. **Model Allowlist** - Specify which models are permitted 1. This will implicitly specify the models that are not permitted, which are any models not included in the allowlist. 2. **Default Model** - Set organization default when multiple allowed 3. **Product scope** - Applied at the instance level (for GitLab Self-Managed and Dedicated) and top level group (for GitLab.com) 1. A subsequent iteration could extend model selection to more granular group level controls. 4. **Feature scope** - Available for Agentic Chat, for the initial iteration. 1. The admin can optionally specify allowed models for Agentic Chat. When specified, those allowed models are presented to end users using agentic chat in the GitLab web UI and IDE extensions. 2. We can consider extending this to foundational flows in a future iteration. 5. **Configuration UI** - This can be configured within the existing model selection page. 6. **Audit Logging** - Track model configuration changes 7. **Supported product offerings** - Duo Agent Platform on GitLab.com, Dedicated, Self-Managed 1. Do not need to support Self-Hosted models in the initial iteration, and we can consider this in a future iteration. ## References * [Design mockups](https://www.figma.com/design/yZNbpRdE9Px4UvHpFciSgE/AI-settings?node-id=3594-5777&p=f&t=D7fI7TcA46sJP2My-0)