Different tool planning behavior between Responses API and Assistants API?

API
, ,
1

Hi Community,

I’m migrating from the Assistants API to the Agents SDK, which I understand is built on top of the Responses API.

I’ve observed that assistants behave differently when calling tools via the Responses API compared to the Assistants API, even when using the same assistant definition, instructions, model settings and tools.

In my case:

  • With the Assistants API, the assistant readily triggers tools based on user intent, even when the prompt is loosely phrased.

  • With the Responses API (tool_choice="auto"), the assistant does not invoke tools unless the prompt is much more explicit.

This leads me to wonder:

  • Does the Assistants API implement any additional tool planning or orchestration beyond what tool_choice="auto" provides?

  • Is this difference in tool behavior expected?

  • Is there a recommended way to replicate the Assistants API’s behavior using the Responses API?

I’d appreciate any clarification from the OpenAI team or anyone who has looked into this. Thanks!

2 Likes