How to force a tool call in a run?

[simone-gaiarin]

Question

How can I force the use of a tool in a specific run of the Agent?
(or is there a better approach than forcing a tool call?)

As an example in LiteLLM it is possible to pass tool_choice to completion.

Additional Context

No response

[DouweM]

@simone-gaiarin The way in PydanticAI to force the model to take a specific action to end the run is using output_type: https://ai.pydantic.dev/output/. Right now, that can only be a type, which could be a Pydantic model that you'll use as input for a function afterwards. In #1785, we'll also be supporting output_type=some_function to force the model to return arguments for that function, after which PydanticAI will call it, and the result of (await agent.run(...)).output will be the function's return value. Would that work for your use case?

[simone-gaiarin]

Thanks, that makes sense and looks like a clean way to force tool usage.

One thing I was trying to reconcile is the recommendation to instantiate agents once and reuse them (e.g., as module-level singletons). That makes sense for general use, especially when the agent has access to multiple tools and decides what to use based on context.

However, in chat-style apps, it might be useful to let users explicitly force a specific tool—for example, "search the web for this." With LiteLLM, this is straightforward since completions are stateless and you can pass tool_choice per call.

Initially, I thought PydanticAI required setting output_type only at agent creation, which would mean instantiating a new agent to force a tool. But after looking at the code, it seems output_type can actually be passed to agent.run(...), which would solve the problem for local tools.

Could you confirm that’s the intended usage?

Also, how would this work when using tools provided by an MCP server? Is there a way to force the use of a specific MCP tool per run, or would that still require configuring the agent differently up front?

[DouweM]

However, in chat-style apps, it might be useful to let users explicitly force a specific tool—for example, "search the web for this." With LiteLLM, this is straightforward since completions are stateless and you can pass tool_choice per call.

@simone-gaiarin That sounds more like #1820, as you'll want the model to use the search results. With output_type, the agent's run would end when the tool is called, and only your code will get to see the result.

I agree it would make sense to also allow that feature on the run call, not just at the agent level. We'd just have to figure out how to indicate only require a tool call on the first request, versus in every request or specific later requests.

That feature would also work with MCP calls.

---

If that sounds more like what you're looking for, I suggest you close this issue and leave some thoughts there!

[simone-gaiarin]

Hi, #1820 is exactly what I was asking. I'll reply there.