AutoGen v0.4 — actor model runtime with fault isolation
runtime — event loop, inbox per actor, delivery with failure isolation
actor: checklist
private state (results, consensus)
inbox deque
receive(message, runtime)
send() puts into runtime queue
no shared memory with other actors
actor: reviewer
private state (verdicts)
handles topic=review
replies with review_result
handler may raise routed to DLQ
other actors keep running
shared runtime queue
m001 checklist "start" [snippets]
m002 reviewer "crash_me" (will fail)
m003-m005 reviewer "review" 3x
worked trace — fault isolation
runtime delivers m002 reviewer raises
DLQ: m002 "RuntimeError: simulated handler failure"
runtime continues delivering m003-m005
reviewer replies review_result for each
checklist reaches consensus across 3 results
three API layers in v0.4
Core AgentRuntime, Agent, Message, Topic
low-level, async, event-driven
AgentChat RoundRobin, Selector, AssistantAgent
task-driven high-level API (v0.2 replacement)
Extensions OpenAI, Anthropic, Azure, tools
Magentic-One is the reference team built on AgentChat
OTel GenAI spans on by default (Lesson 23)
v0.7.x in maintenance. Microsoft Agent Framework (preview Oct 2025) is the forward path — same ideas.