Support managed agents in ToolCallingAgent #1456
Conversation
There was a problem hiding this comment.
Pull Request Overview
This PR adds support for managed agents in ToolCallingAgent, ensuring they are treated like tools when calling the model and surfaced in prompts.
- Register managed agents as callable tools with input/output metadata
- Update prompt templates to display each agent’s inputs and output type
- Include managed agents in
tools_to_call_fromfor both streaming and non-streaming calls
Reviewed Changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| tests/test_agents.py | Added a test to verify that managed agents are passed to model.generate alongside tools |
| src/smolagents/prompts/toolcalling_agent.yaml | Extended system prompt to list agent inputs and output types |
| src/smolagents/prompts/structured_code_agent.yaml | Updated example agent signatures to include typed task argument and docstring |
| src/smolagents/prompts/code_agent.yaml | Mirrored signature/docs update in code agent prompt |
| src/smolagents/agents.py | In _setup_managed_agents, set inputs/output_type; merged managed agents into tools_to_call_from |
Comments suppressed due to low confidence (2)
tests/test_agents.py:1089
- Consider adding a complementary test for the non-streaming path (when
stream=False) to verify that managed agents are also included inmodel.generatecalls in that scenario.
def test_toolcalling_agent_passes_both_tools_and_managed_agents(self, test_tool):
src/smolagents/prompts/toolcalling_agent.yaml:106
- [nitpick] The indentation of these lines is deeper than the preceding list item. Align them with the bullet (
-) for consistent prompt formatting.
Takes inputs: {{agent.inputs}}
| @@ -150,8 +150,12 @@ system_prompt: |- | |||
| Here is a list of the team members that you can call: | |||
| ```python | |||
| {%- for agent in managed_agents.values() %} | |||
| def {{ agent.name }}("Your query goes here.") -> str: | |||
| """{{ agent.description }}""" | |||
| def {{ agent.name }}(task: str) -> str: | |||
There was a problem hiding this comment.
You can add additional_args here to hint the manager that it can pass other args:
| def {{ agent.name }}(task: str) -> str: | |
| def {{ agent.name }}(task: str, additional_args: dict[str, Any]) -> str: |
There was a problem hiding this comment.
I guess this is a follow-up to:
I also guess this should be added to the ToolCallingAgent prompt as well...
There was a problem hiding this comment.
I think I will treat that in a subsequent PR.
|
|
||
| Args: | ||
| task: Long detailed description of the task. | ||
| """ |
There was a problem hiding this comment.
| """ | |
| additional_args: a dictionary of other variables to make available to your managed agent, e.g. images, or dataframes. | |
| """ |
There was a problem hiding this comment.
I just noted that the images param is treated differently from the additional_args param in the run method.
Is this expected? I mean, if we can to pass images to the managed agent, should we use images or additional_args?
There was a problem hiding this comment.
I think I will treat that in a subsequent PR.
Support managed agents in
ToolCallingAgent.Fix #1455.
Fix #1353.
Supersede and close #1418.