[FEATURE] Add better coworkers context for crews with managers

[Saicheg]

Feature Area

Agent capabilities

Is your feature request related to a an existing bug? Please link it here.

NA

Describe the solution you'd like

Hi there!

Before making a suggestion just let me take a quick work and thank you for this amazing framework. I fall in love with simplicity and power of it a minute I found it.

Lately I am wrapping my head around crews with managers. Recently I discovered that the way "management" is done is just by assigning two tools in order to allow manager to delegate a work.

Tool Name: Delegate work to coworker
Tool Arguments: {'task': {'description': 'The task to delegate', 'type': 'str'}, 'context': {'description': 'The context for the task', 'type': 'str'}, 'coworker': {'description': 'The role/name of the coworker to delegate to', 'type': 'str'}}
Tool Description: Delegate a specific task to one of the following coworkers: News and Trend Researcher, Assistant, Topic Researcher
The input to this tool should be the coworker, the task you want them to do, and ALL necessary context to execute the task, they know nothing about the task, so share absolute everything you know, don't reference things but instead explain them.

Tool Name: Ask question to coworker
Tool Arguments: {'question': {'description': 'The question to ask', 'type': 'str'}, 'context': {'description': 'The context for the question', 'type': 'str'}, 'coworker': {'description': 'The role/name of the coworker to ask', 'type': 'str'}}
Tool Description: Ask a specific question to one of the following coworkers: News and Trend Researcher, Assistant, Topic Researcher

Problem for me was that in current implementation default manager can only understand which coworker to ask by their name. There is no additional context about coworkers goals/descriptions which will affect manager description. This means if you have 3 agents ( Researcher, Quick Researcher, Deep Researcher ) for example - manager will be very confused.

I would love to have a way to describe coworkers to manager with better context other that just name. It could be by including their descriptions /goals or having additional field for manager like manager_notes:.

Describe alternatives you've considered

The way I am solving this right now is by creating fully-custom manager with JSON description for each of my agents. But it is a little messy and I believe should be something by default in framework:

    @property
    def manager(self) -> Agent:
        return Agent(
            allow_delegation=True,
            role="Crew Manager",
            goal="Manage the team to complete the task in the best way possible.",
            backstory=dedent(
                """
                You are a seasoned manager with a knack for getting the best out of your team.
                You are also known for your ability to delegate work to the right coworkers, and to ask the right questions to get the best out of your team.
                Even though you don't perform tasks by yourself, you have a lot of experience in the field, which allows you to properly evaluate the work of your team members.

                Below is JSON description for your coworkwers abilities:
                {coworkers_prompt}
                """
            ).format(coworkers_prompt=self.coworkers_prompt).strip(),
        )

Additional context

No response

Willingness to Contribute

Yes, I'd be happy to submit a pull request

[Vidit-Ostwal]

@Saicheg, I think this is a good idea, to include some context, idea about the agents or tools the manager llm can use.
The only problem I see is that internally manager llm is just like an agent only, and therefore adding a manager_note kind of thing just for one agent, might not be the best idea.

Another thing I can see is that one can ask an agent to make description, goal, context for manager agent,
basically make the agent, if not provided and use the context of different agents and tools available.

[Saicheg]

Another thing I can see is that one can ask an agent to make description, goal, context for manager agent, basically make the agent, if not provided and use the context of different agents and tools available.

@Vidit-Ostwal thanks for your feedback! I like the idea of creating such description on fly, but I am a little concerned about a delay ( since it will obviously use LLM and API request for that ) and tokens cost.

Easiest solution will be to use goal + tools descriptions for agent agent glued someway, maybe even JSON?

[Vidit-Ostwal]

Yes, pre-computed or may be a generalized manager, but I think better to have that kind of control with you.

[bhancockio]

Hey @Saicheg!

So there are 2 ways to create a manager agent right now.

The first is to set the manager_llm property on your crew. Inside the app, we create a new manager_agent that comes with some existing instructions (check out the first few lines of the en.json file to see the boilerplate agent instructions).

The second approach is to create your own manager_agent and set that.

Based on what you were doing earlier, it looks like you've already created your own custom manager agent so I would recommend going with the second approach.

Ultimately, you'll end up with something like this:

custom_manager = Agent(....) # <-- Put your custom instructions here

crew(
  agents=...
  tasks=...
  manager_agent=custom_manager

Please let me know if you have any other questions!

For now, I'm going to close this issue.

[Saicheg]

@bhancockio thanks for your attention!

I actually have a question if you don't mind.

I think that default behavior for manager_llm should be different and at least include each agent role as part of default prompt. I am not saying that there is no other way around to solve this problem ( aka with manager_agent ), but I was pretty surprised to find that, by default, it is just choosing agents based on their names.

Don't you think that default behavior should be different if you use manager_llm?

If you think that it makes sense, I am happy to contribute! I just want a green light from some of the core contributors.