Instructions

You are creating an AG2 nested chat workflow. This encapsulates a multi-step pipeline inside a single agent using register_nested_chats.

Instructions

1. Ask the user for:

   • The pipeline stages and what each agent does
   • The trigger condition (which sender activates the nested pipeline)
   • How results pass between stages (summary_method)

2. Create the nested chat following this pattern:

Nested Chat Pattern

import asyncio
from autogen import ConversableAgent, LLMConfig

llm_config = LLMConfig({"api_type": "anthropic", "model": "claude-sonnet-4-6"})

user = ConversableAgent(
    name="user",
    human_input_mode="NEVER",
)

# The outer agent that encapsulates the pipeline
coordinator = ConversableAgent(
    name="coordinator",
    system_message="Present the final result.",
    llm_config=llm_config,
)

# Pipeline stage agents
step_1 = ConversableAgent(
    name="step_1",
    system_message="Do the first step.",
    llm_config=llm_config,
)

step_2 = ConversableAgent(
    name="step_2",
    system_message="Do the second step.",
    llm_config=llm_config,
)

step_3 = ConversableAgent(
    name="step_3",
    system_message="Do the third step.",
    llm_config=llm_config,
)

# Register the nested pipeline -- fires when coordinator receives from user
coordinator.register_nested_chats(
    chat_queue=[
        {
            "recipient": step_1,
            "message": lambda recipient, messages, sender, config: messages[-1]["content"],
            "max_turns": 1,
            "summary_method": "last_msg",
        },
        {
            "recipient": step_2,
            "message": "Continue with the second step.",
            "max_turns": 1,
            "summary_method": "last_msg",
        },
        {
            "recipient": step_3,
            "message": "Complete the third step.",
            "max_turns": 1,
            "summary_method": "last_msg",
        },
    ],
    trigger=user,  # fires when message comes from user
)


async def main():
    response = await user.a_run(
        coordinator,
        message="Your task here",
        max_turns=1,
    )
    await response.process()
    print(await response.summary)


if __name__ == "__main__":
    asyncio.run(main())

Key Rules

• Use register_nested_chats to define the pipeline on the outer agent
• Each chat in chat_queue MUST have a message field -- without it, subsequent stages won't fire
• The first stage should use a callable message to forward the original user request: lambda recipient, messages, sender, config: messages[-1]["content"]
• Subsequent stages can use static message strings -- the previous stage's output is automatically appended as context
• Each chat in chat_queue runs sequentially -- output of one feeds into the next
• Use max_turns=1 per stage for clean handoffs
• summary_method="last_msg" passes the last message as input to the next stage
• The trigger parameter controls which sender activates the nested pipeline
• Use a_run (async) with .process() then .summary -- NOT initiate_chat
• Use LLMConfig({...}) -- NOT a raw dict like {"model": "..."}

When to Use This Pattern

• Packaging a complex workflow into a single agent interface
• When the pipeline should be invisible to the caller
• Multi-step processing: research -> draft -> review -> polish
• Hub-and-spoke where a coordinator consults multiple specialists

Example

See examples/article_pipeline.py for a research-draft-edit pipeline.