Access tool call items in an output guardrail

[jonjomckay]

Is it possible to check which tools were called, and their inputs/outputs in a guardrail? My use case is as follows:

1. The user creates an "agent" entity in our platform, configured to look at person-1 and person-2
2. Our platform runs an Agent using this SDK, telling it to use the queryPerson tool through MCP with the IDs person-1 and person-2
3. We receive results for both people

I'd like to ensure that the agent did in fact call the queryPerson tool, with person-1 and person-2 as parameters. This does happen 99% of the time, as we're explicitly telling the LLM to, but there are occasions where it doesn't and it seems like an output guardrail might be the appropriate place to catch this (and e.g. retry the agent, or return an error).

I have stepped through the parameters passed to my OutputGuardrail.execute function, but I can't seem to find any of the tool calls. They do appear in the result from the run function though, but it's too late to be used in a guardrail by then.

[seratch]

Thanks for asking the question. You can use details?.modelResponse?.output in the output guardrail argument list for it. You can check the function tool call items to verify things go as you expect.

[jonjomckay]

Thanks for the pointer! I just double-checked and details.modelResponse.output only has a single message - my agent's structured output, and doesn't include any tool calls.

The RunResult returned from my run call does contain a whole lot more data though, including history and output, which do contain all the tool calls. Would it be difficult to expose these (or the whole RunResult through the output guardrail? It looks like they currently only support guarding on the final output message, rather than the whole agent flow -> output.

[seratch]

Thanks for your reply, and you're right that there is currently no way to access the tool call items! My pull request #716 will resolve this issue. We expect releasing a new patch version including this change soon (by the end of this week for sure).

[jonjomckay]

Awesome, thank you!