Issue Updating Agent with Non-ChatGPT Models using Assistant API

Question

I’m experiencing an issue while working with Azure AI Foundry agents.

I’m able to create agents using non-ChatGPT models (for example, Llama-4-Maverick-17B-128E-Instruct-FP8). However, when I try to update an existing agent to use one of these models, the request fails and returns an error.

Additionally, when using the Assistant API, I can successfully create an agent with the Llama-4-Maverick-17B-128E-Instruct-FP8 model, but I’m unable to update the agent to use this model afterward.

Answer 1

Updating an existing agent to use a different model is only supported where the underlying API/SDK allows modifying the agent’s model property on update. In the Azure AI Foundry ecosystem there are two relevant patterns:

1. New Foundry Agents (recommended)
   For Foundry Agents, model changes are typically handled by creating a new agent version rather than updating the model in-place. The migration and troubleshooting guidance notes that older methods like create_agent() were removed and replaced with create_version() / CreateAgentVersionAsync() / createAgentVersion() that take a PromptAgentDefinition including the model:
   • To “change” the model for an agent, create a new agent version with the desired model in the PromptAgentDefinition instead of trying to update the existing version’s model.
   • If the new model performs poorly or causes issues, use the Agent Playground version history to roll back to a previous version. The rollback restores the earlier model configuration for that agent.
   This versioned approach is the supported way to move an agent between models in Foundry, including non-ChatGPT models, and avoids relying on an in-place model update that may not be supported.
2. Persistent/Assistants-style APIs
   For the .NET Persistent Agents client (which follows the Assistants semantics), the PersistentAgentsAdministrationClient.UpdateAgentAsync method has a model parameter specifically documented as:

   The ID of the model to use.

   This means, in principle, updating an agent’s model is supported by passing the new model ID into UpdateAgentAsync along with the assistantId. However, whether a particular non-ChatGPT model (such as Llama-4-Maverick-17B-128E-Instruct-FP8) is accepted depends on:
   • The model being valid and enabled for the Assistants/Agents feature in the region and resource.
   • The service-side constraints for which models can be used with Assistants/Agents. Not all deployed models are necessarily valid for Assistants/Agents, even if they can be used for plain completions.
   If creation succeeds but updating fails for the same model, this typically indicates a service limitation or bug rather than a client-side misuse, because the same model ID is already accepted at creation time.
3. Assistants API deprecation and Foundry Agents
   The Assistants API is explicitly marked as deprecated and scheduled for retirement, with guidance to move to the Microsoft Foundry Agents service. For long-term and fully supported behavior (including model changes), the recommended path is:
   • Migrate workloads from Assistants to Foundry Agents.
   • Use agent versioning (create a new version with the new model) instead of relying on in-place model changes.
4. Practical mitigation for the described issue Given that:
   • Agent creation with the non-ChatGPT model works, but
   • Updating the agent to use that model fails,
   the supported and reliable pattern in Foundry is:
   • Treat model changes as new versions of the agent rather than updates to the existing one.
   • For Foundry Agents: call the appropriate create_version/CreateAgentVersionAsync/createAgentVersion with a PromptAgentDefinition that references the non-ChatGPT model.
   • Use Agent Playground version history to roll back if needed.
   For Persistent/Assistants-style agents, if UpdateAgentAsync with the model parameter fails specifically for this model while creation works, this is likely a platform issue that requires service-side investigation.

---

References:

• Migrate to the new agents developer experience
• Connect an AI gateway to Foundry Agent Service (preview)
• PersistentAgentsAdministrationClient.UpdateAgentAsync Method
• Azure OpenAI in Microsoft Foundry Models REST API preview reference
• Azure OpenAI in Azure AI Foundry Models
• Upgrade or switch models with Ask AI (preview)