Ag2 Chat Fundamentals

AG2 Chat Fundamentals

These conventions apply to ALL AG2 agent patterns (two-agent chat, sequential chat, nested chat, group chat).

References

• AG2 GitHub: https://github.com/ag2ai/ag2
• AG2 Docs: https://docs.ag2.ai/latest/
• API Reference: https://docs.ag2.ai/latest/docs/api-reference/autogen/Agent/

LLM Configuration

Always use LLMConfig -- never pass a raw dict as llm_config.

from autogen import LLMConfig

# Correct -- LLMConfig wrapping a dict
llm_config = LLMConfig({"api_type": "anthropic", "model": "claude-sonnet-4-6"})

# Wrong -- raw dict
# llm_config = {"model": "claude-sonnet-4-6"}

Agent Naming

Agent names must never contain spaces. Use lowercase with underscores for multi-word names:

# Correct
name="project_manager"
name="qa_engineer"

# Wrong
name="Project Manager"
name="QA Engineer"

Agent description Field

The description parameter is critical for AutoPattern group chats -- the underlying Group Chat Manager uses it to decide which agent to select next. Every agent in an AutoPattern group chat MUST have a distinct description explaining when to select that agent. The system_message tells the agent how to behave; the description tells the manager when to route to that agent.

For non-group-chat patterns (two-agent, sequential, nested), description is optional.

Agent Types

• ConversableAgent: The base agent. Use for most agents. Defaults: human_input_mode="TERMINATE", code_execution_config=False.
• UserProxyAgent: A human-in-the-loop proxy. Defaults: human_input_mode="ALWAYS", code_execution_config={} (enabled). If you want human_input_mode="NEVER" with no code execution, just use ConversableAgent instead.

Ending a Chat

There are several ways agent conversations end in AG2. These apply to two-agent chats, sequential chats, and group chats.

1. Maximum Turns / Rounds

For two-agent chats, use max_turns on a_run:

response = await agent_a.a_run(
    agent_b,
    message="Your task",
    max_turns=3,
)

For group chats, use max_rounds on run_group_chat:

result = run_group_chat(
    pattern=pattern,
    messages="Your task",
    max_rounds=15,
)
result.process()
print(result.summary)

2. TerminateTarget (Group Chat)

An agent hands off to TerminateTarget() to explicitly end a group chat.

from autogen.agentchat.group import TerminateTarget, ReplyResult

# As after-work behavior
final_agent.handoffs.set_after_work(TerminateTarget())

# From a tool via ReplyResult
def finish_task(summary: str, context_variables: ContextVariables) -> ReplyResult:
    """Complete the task and end the conversation."""
    context_variables["final_summary"] = summary
    return ReplyResult(
        message=f"Task complete: {summary}",
        context_variables=context_variables,
        target=TerminateTarget(),
    )

3. RevertToUserTarget (Group Chat)

Returns control to the user agent. If no user_agent is configured in the pattern, this ends the conversation.

from autogen.agentchat.group import RevertToUserTarget

agent.handoffs.set_after_work(RevertToUserTarget())

4. Termination Message

An agent configured with is_termination_msg ends the chat when it receives a matching message.

agent = ConversableAgent(
    name="agent",
    is_termination_msg=lambda msg: "TERMINATE" in (msg.get("content", "") or ""),
    llm_config=llm_config,
)

5. Human Exit

When human_input_mode="ALWAYS" (the UserProxyAgent default), the user can type exit to end the conversation immediately.

6. Max Consecutive Auto Reply

Limits how many times an agent can reply consecutively before stopping.

agent = ConversableAgent(
    name="agent",
    max_consecutive_auto_reply=3,
    llm_config=llm_config,
)

Chat APIs

Use the modern async APIs:

Call .process() to run the workflow, then access .summary for the result:

response = await agent_a.a_run(agent_b, message="Hello", max_turns=2)
await response.process()        # processes events, returns None
print(await response.summary)   # async property -- must await

Summary Methods

When passing output between stages, use summary_method:

• "last_msg" -- pass the last message as-is (default for pipelines)
• "reflection_with_llm" -- use an LLM to summarize the conversation