AutoGen (Microsoft) enables multi-agent conversations where agents collaborate by exchanging messages. AssistantAgent wraps an LLM with a system prompt and tools. UserProxyAgent executes code and relays human input. GroupChat orchestrates multiple agents in a shared conversation. The code execution sandbox (Docker or local subprocess) lets agents write and run code iteratively. AutoGen 0.4 introduced an actor-model-based runtime (AgentChat) for improved scalability.
Each stage in order — click any step to read what it does.
Key AutoGen agent types and their roles.
The trade-offs worth knowing before you build this.
LocalCommandLineCodeExecutor runs code on the host — a security risk. DockerCommandLineCodeExecutor isolates execution in a container with no network or filesystem access.
Without a termination condition, GroupChat can loop indefinitely. The is_termination_msg function (checks for "TERMINATE") or max_turns parameter must be set on every agent.
Each agent in a GroupChat should have a distinct persona in its system_message. Without clear role boundaries, agents repeat each other's work or talk past each other.
Sign in to share your feedback and join the discussion.
