Specifying `model_provider` for each `Agent`

[qqaatw]

Hello,

For now, the model_provider argument can only be passed to Runner.run, which is alright in cases where all the handoff agents use the same api_key and api_version.

In our use cases, we have different api_keys and api_versions for gpt and reasoning models, and we want to combine these models to build a multi-agent system with agent handoffs.

Ideally, we would like to be able to provide different model providers for each agent rather than a global model provider. Any input is appreciated.

[seratch]

Thanks for sharing this feedback. Our current recommendation is to use different runners for agents that require different api keys, but we may consider some enhancement in the future. At this moment, there is nothing to share and we don't have any immediate plans.

[qqaatw]

I see. If runners are different, question would be - how do handoffs work?

[seratch]

I know this is not the elegant solution you’d like to see, but having another agent (or logic without LLM) to decide which agent to call next, and then running that agent as the next step, is a possible approach.

Also, if you don’t need to stick to the handoff design pattern, another solution would be to use a different API key or model agent as an internal execution within a function tool, and then call that function tool as a tool. This way, you can simply instantiate the agent and runner within the function, or for the function, as needed.

Hope this helps.

[qqaatw]

I see, thank you @seratch

[seratch]

Let me close this for now. If you have a different topic/question, please feel free to create a new issue for it.