Swarm: User-defined AfterWork string for agent selection using LLM

[marklysze]

Why are these changes needed?

In line with #26, this provides the developer with an option to specify a string prompt for an LLM to use to select the next agent in an agent's AfterWork hand off.

It extends the AfterWorkOption.SWARM_MANAGER (which, by default, uses the default auto speaker selection options) with an additional AfterWork parameter, next_agent_selection_msg, which becomes the selection prompt used by the LLM.

next_agent_selection_msg can be a string or a Callable that returns a string.

The signature for the Callable is:
def my_selection_message(agent: ConversableAgent, messages: List[Dict[str, Any]]) -> str

The string provided, or returned from the Callable, will have the following string substitutions:

• {agentlist} will be replaced by a comma-delimited list of agents
• context variable keys will be replaced by their values

The underlying groupchat's GroupChatManager will be used for the speaker selection and must, therefore, have an LLM Config. This can be set through the new swarm_manager_args parameter on a/initiate_swarm_chat.

Example:

context_variables = {
    "order_id": "ABC1234"
}

register_handoff(
    from=planner,
    hand_to=[
        AfterWork(agent=AfterWorkOption.SWARM_MANAGER, next_agent_selection_msg="Select the next agent from this list {agentlist}. From the list please select the order_mgmt_agent. Order Id is {order_id}"),
    ]
)

_, _, _ = initiate_swarm_chat(
    ...,
    swarm_manager_args={"llm_config": llm_config}
)

Related issue number

#26

Checks

• I've included any doc changes needed for https://docs.ag2.ai/. See https://docs.ag2.ai/docs/contributor-guide/documentation to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.

[marklysze]

Thanks for the initial review @sonichi, if you could have another look over now that would be appreciated.

[marklysze]

@yiranwu0, @linmou - if you're able to test this it would be appreciated!

[yiranwu0]

Look good to me overall. My only question is about the context_str class. I think this class might not be necessary at all, and also consider existing packages like dedent.

[marklysze]

Thanks all, merging so we can merge swarm to conversable.