Table of Contents
🌏 中文版
AG2 is an open-source Python framework for agents and multi-agent workflows. Its central abstraction is not a graph node or fixed task but a conversable agent: each role can receive messages, call a model, run tools, answer another agent, or wait for human input.
Developers familiar with early AutoGen will recognize the autogen import and ConversableAgent. The installation guide states that autogen and ag2 are aliases for the same PyPI package. This is also the main naming trap: AG2 and Microsoft's later Microsoft Agent Framework are separate projects.
ConversableAgent is the common base
The official guide identifies ConversableAgent as AG2's fundamental class. A name, system message, and model configuration establish the role; tools, human input, termination rules, and reply functions extend it.
from autogen import ConversableAgent, LLMConfig
llm_config = LLMConfig({"api_type": "openai", "model": "gpt-5-nano"})
reviewer = ConversableAgent(
name="reviewer",
system_message="Review the draft and return concrete corrections.",
llm_config=llm_config,
)
result = reviewer.run(message="Review this release note.", max_turns=2)
result.process()run() returns an event iterator rather than finished text. Applications can consume events for a UI, logging, or approvals; process() is a console helper. Production code should therefore inspect events and outcomes instead of treating printed output as success.
GroupChat turns speaker choice into orchestration
With more than two roles, GroupChat keeps a shared conversation thread and lets GroupChatManager choose the next speaker. Built-in strategies include round robin, random, manual, and LLM-selected routing. Transition constraints can prevent arbitrary handoffs.
This fits researcher-writer-reviewer collaboration because “who should act next?” is the central uncertainty. The tradeoff is less predictable behavior and cost. Automatic speaker selection is itself a model decision, while shared history keeps expanding. Put hard limits on rounds, tools, and termination conditions, then trace why each turn occurred.
Conversation history is not a business checkpoint
AG2 records who said what, but a message list does not automatically become replayable order, approval, or pipeline state. External side effects still need idempotency keys, and approvals lasting days require a database or durable runtime.
AG2 is therefore a strong fit for role-based prototypes, human-in-the-loop conversations, and existing AutoGen codebases. For fixed ETL, payment, or precisely resumable workflows, keep authoritative state elsewhere and use AG2 only where language-model judgment is needed.
Overall
AG2 makes multi-agent coordination read like a conversation instead of a state machine. Start with two roles and round robin, then compare success rate, token cost, and latency against a single-agent baseline. Add automatic speaker selection only when role separation produces measurable value. For the wider landscape, see the agent framework selection guide.
References
Loading...