Agent as tool

[sordonia]

• adds agent as tool capability

[Copilot]

Pull Request Overview

This PR adds agent-as-tool support by exposing tool choice in the OpenAI wrapper, refactoring LLM instantiation pathways, and introducing a new AgentTool.

• Allow passing tool_choice via kwargs in openai.py
• Split and rename LLM instantiation into instantiate_from_config and a new instantiate in base.py
• Add AgentTool for querying a specialized LLM agent

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
debug_gym/llms/openai.py	Make tool_choice configurable via kwargs.pop
debug_gym/llms/base.py	Refactor instantiate methods, rename one to instantiate_from_config
debug_gym/gym/tools/agent.py	New AgentTool class for agent-as-tool capability

Comments suppressed due to low confidence (2)

debug_gym/llms/openai.py:185

• [nitpick] The docstring for generate should be updated to document the new tool_choice parameter so users know they can override it via kwargs.

    def generate(self, messages, tools, **kwargs) -> LLMResponse:

debug_gym/gym/tools/agent.py:13

• [nitpick] No unit tests were added for AgentTool. Consider adding tests for use, __copy__, and __deepcopy__ behaviors to ensure correct functionality.

@Toolbox.register()

[xingdi-eric-yuan]

I feel the agentic_tool can be something more generic, and this history summarization tool being an instantiation of such class?

[sordonia]

I feel the agentic_tool can be something more generic, and this history summarization tool being an instantiation of such class?

what do you have in mind? @xingdi-eric-yuan there's no history summarization here

[xingdi-eric-yuan]

Might be useful to have some documentation/instruction on how to use this tool? But it can also be in a future PR.

[MarcCote]

Left a comment to be addressed before merging.

[MarcCote]

What is this history? A reference to an agent's history?

[sordonia]

it's the HistoryTracker of the caller agent, the agentic tool might want to peek at the interactions of the caller agent

[MarcCote]

But do you have an example of how it is use?