Alui/multiagent

[luiarthur]

No description provided.

[luiarthur]

@mikegros This draft demonstrates how to include the planning and execution agents in a supervisor agent (that I call Ursa).

Changes:

• src/agents/excution_agent.py
  • minor change to aid type checking
• src/experimental/agents/multiagent.py contains
  • a system prompt for the supervisor
  • tool wrapped execution and planning agents
  • an additional tool that handles splitting plan steps into different execution agent calls
• tests/agents/test_multiagent/test_multiagent.py
  • a test/demo

A good way to start reviewing the changes would be to pull the branch and then try to run test_multiagent.py by running from the root directory

pytest -s tests/agents/test_multiagent

[mikegros]

I think this looks great overall. A couple things:

• I like having a main agent names Ursa
• I think the type checking in the execute agent may be a little strict. I saw that you needed extra_tools to be a list of BaseTools - I assume that a StructuredTool counts as a BaseTool, but the add_tools capability also can take just plain functions, which is part of why the typing was originally a Callable. I think that was wrong (that we wanted it to be a Callable or some sort of tool object), but I think requiring just a BaseAgent might not be ideal (that said, extra_tools might be a weird thing to have anyway since we have the add tools method.

[mikegros]

See my comment in the PR

[luiarthur]

@mikegros Thanks for reviewing! My main hesitation with this PR is I don't know if the memory is being properly handled. In this implementation, the Ursa agent retains memory; no other agents retain memory. Since the outputs of subagents are returned to the Ursa agent, I would assume there's no need to pass around the history to other agents. Does that make sense?

[mikegros]

@mikegros Thanks for reviewing! My main hesitation with this PR is I don't know if the memory is being properly handled. In this implementation, the Ursa agent retains memory; no other agents retain memory. Since the outputs of subagents are returned to the Ursa agent, I would assume there's no need to pass around the history to other agents. Does that make sense?

I will look at that closer, but as a quick response, you could give all the tool agents the same checkpointer database, as long as they all have their own thread_id (which is the way the CLI handles checkpointing). Then they could all keep a message history.

[mikegros]

I didnt realize I had filled out this review but not submitted it. I think this is mostly good to merge. Especially since it is listed as "experimental" so if memory aspects aren't being handled ideally, that can be part of the iteration that happens between now and when we would make it no-longer-experimental.

[mikegros]

Should this inherit BaseAgent for usage metrics or anything else?

[mikegros]

We talked about this, so I dont think we need to make any changes.