Run Tools Sequentially

[strawgate]

Description

There was previously an issue opened here #462 for running tools sequentially, which suggested disabling parallel_tool_calls on the model.

This forces the model to do one tool call per turn which significantly inflates the turn count and cost of running the model as each tool call is now a round trip.

Ideally there would be the option to allow the Agent to return multiple tool calls but to perform the actual tool calls sequentially, in the order they were provided by the Agent.

References

No response

[pydanticai-bot]

PydanticAI Github Bot Found 1 issues similar to this one:

1. "Question - Can agent call multiple tools in one go? #1096" (95% similar)

[strawgate]

Im wondering if it might make sense to have this be a per-toolset setting where we put a semaphore on the toolset itself? Maybe as a wrapper toolset?

[DouweM]

@strawgate I like doing this with a special toolset, possibly with a contextvar to hold the semaphore so it'll work per-agent when the toolset is used by multiple agents in parallel (and the context the contextvar lives inside is the one created in agent.iter). Feel free to submit a PR!

[strawgate]

A semaphore allows running tools one at a time but doesn't actually offer running them sequentially.

Ideally we'd be able to offer the ability for tool calls to occur in the order they are provided by the LLM.

For example, if the LLM is performing Git operations, it's very important that git add runs before git commit, etc. Or when performing filesystme operations, updating a file and then reading it.

I dont think there's anything (reasonable) we can do in a toolset to order the requests.

In my testing it's somewhat difficult to convince the model to run tool calls in separate steps but very easy to convince the model to run tool calls in a specific order (it seems to run them in the right order automatically)

[DouweM]

@strawgate Looks like an argument on Agent and/or agent.run (etc) would be the only option then. Feel free to submit a PR.

[strawgate]

@DouweM starting to feel like we need an AgentSettings...

[strawgate]

I guess an alternative option would be to indicate that the tool order matters on the tool definition and then check if any of the called tools have allow_parallel: False (or whatever) and if any of the called tools have that set, then tool calls are performed sequentially.

will try a PR to put this on Agent or run

[strawgate]

i think to do this i'd need to move this through the graph context and the Agent and modify all the run/iter/run_stream in abstract agent, etc

How would you feel about having the allow_parallel = True as the default on ToolDefinition and then having a method on ToolManager called like should_parallel_tool_call which takes a list of desired tool calls and returns whether they should run in parallel (checking if every tool has allow_parallel=true)

[strawgate]

@DouweM added an example PR -- I imagine if this was agreeable, adding allow_parallel to the tool decorator on Agent (and to the function toolset, etc) would be the next step