AI 工程从零开始

主题

AutoGen v0.4: Actor Model and Agent Framework ​

AutoGen v0.4 (Microsoft Research, Jan 2025) redesigned agent orchestration around the actor model. Async message exchange, event-driven agents, fault isolation, natural concurrency. The framework is now in maintenance mode while Microsoft Agent Framework (public preview Oct 2025) becomes the successor.

Type: Learn + Build Languages: Python (stdlib) Prerequisites: Phase 14 · 01 (Agent Loop), Phase 14 · 12 (Workflow Patterns) Time: ~75 minutes

Learning Objectives ​

  • Describe the actor model: agents as actors, messages as the only IPC, failure isolation per actor.
  • Name AutoGen v0.4's three API layers — Core, AgentChat, Extensions — and what each is for.
  • Explain why decoupling message delivery from handling gives fault isolation and natural concurrency.
  • Implement a stdlib actor runtime in Python and port a two-agent code-review flow onto it.

The Problem ​

Most agent frameworks are synchronous: one agent produces, one agent consumes, in a call stack. Failures crash the stack. Concurrency is bolted on. Distribution requires rewriting.

AutoGen v0.4's answer: the actor model. Each agent is an actor with a private inbox. Messages are the only interaction. The runtime decouples delivery from handling. Failures isolate to one actor. Concurrency is native. Distribution is just different transport.

The Concept ​

Actors ​

An actor has:

  • A private state (never directly touched from outside).
  • An inbox (message queue).
  • A handler: receive(message) -> effects where effects can be "reply," "send to other actor," "spawn new actor," "update state," "stop self."

Two actors cannot share memory. They can only send messages.

Three API layers in AutoGen v0.4 ​

  1. Core. Low-level actor framework. AgentRuntime, Agent, Message, Topic. Async message exchange, event-driven.
  2. AgentChat. Task-driven high-level API (replacement for v0.2's ConversableAgent). AssistantAgent, UserProxyAgent, RoundRobinGroupChat, SelectorGroupChat.
  3. Extensions. Integrations — OpenAI, Anthropic, Azure, tools, memory.

Why decoupling matters ​

In the v0.2 model, calling agent_a.chat(agent_b) synchronously blocks agent_a until agent_b returns. In v0.4, send(agent_b, msg) puts the message in agent_b's inbox and returns. The runtime delivers later. Three consequences:

  • Fault isolation. Agent B crashing does not crash Agent A — the runtime catches the failure in B's handler and decides what to do (log, retry, dead-letter).
  • Natural concurrency. Many messages in flight at once; actors process their inbox concurrently.
  • Distribution-ready. Inbox + transport is the same abstraction whether the actor is in-process or on another host.

Topologies ​

  • RoundRobinGroupChat. Agents take turns in a fixed rotation.
  • SelectorGroupChat. A selector agent picks who goes next based on conversation context.
  • Magentic-One. Reference multi-agent team for web browsing, code execution, file handling. Built on AgentChat.

Observability ​

OpenTelemetry support is built in. Every message emits a span; tool calls carry gen_ai.* attributes per the 2026 OTel GenAI semantic conventions (Lesson 23).

Status: maintenance mode ​

Early 2026: AutoGen v0.7.x is stable for research and prototyping. Microsoft has shifted active development to the Microsoft Agent Framework (public preview Oct 1 2025; 1.0 GA targeted end of Q1 2026). AutoGen patterns port forward cleanly — the actor model is the durable idea.

Build It ​

code/main.py implements a stdlib actor runtime:

  • Message — typed payload with sender, recipient, topic, body.
  • Actor — abstract with receive(message, runtime).
  • Runtime — event loop with a shared queue, delivery, failure isolation.
  • A two-actor demo: ReviewerAgent reviews code, ChecklistAgent runs a checklist; they exchange messages until consensus.

Run it:

python3 code/main.py

The trace shows message delivery, a simulated failure in one actor that does not crash the other, and convergence on a shared verdict.

Use It ​

  • AutoGen v0.4/v0.7 (maintenance) — stable for research, prototyping, multi-agent patterns.
  • Microsoft Agent Framework (public preview) — the forward path; same actor-model ideas in a refreshed API.
  • LangGraph swarm topology (Lesson 13) — similar pattern via shared-tool handoffs.
  • Custom actor runtime — when you need specific transport (NATS, RabbitMQ, gRPC).

Ship It ​

outputs/skill-actor-runtime.md generates a minimal actor runtime plus a team template (RoundRobin or Selector) for a given multi-agent task.

Exercises ​

  1. Add a dead-letter queue: when a handler raises, park the failing message for human inspection. How often does DLQ get hit in your toy?
  2. Implement SelectorGroupChat: a selector actor picks who processes the next message based on conversation state.
  3. Add distributed transport: swap the in-process queue for a JSON-over-HTTP server so actors can run in separate processes.
  4. Wire an OTel span per message (or a no-op stand-in). Emit gen_ai.agent.name, gen_ai.operation.name per Lesson 23.
  5. Read AutoGen v0.4's architecture post. Port your toy to the real autogen_core API. What did you skip that matters in production?

Key Terms ​

TermWhat people sayWhat it actually means
Actor"Agent"Private state + inbox + handler; no shared memory
Message"Event"Typed payload; the only way actors interact
Inbox"Mailbox"Per-actor queue of pending messages
Runtime"Agent host"Event loop that routes messages and isolates failures
Topic"Channel"Named publish-subscribe route between actors
Fault isolation"Let it crash"One actor failing does not crash others
RoundRobinGroupChat"Fixed-rotation team"Agents take turns in order
SelectorGroupChat"Context-routed team"Selector picks who goes next
Magentic-One"Reference team"Multi-agent squad for web + code + files

Further Reading ​