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,
)