AutoGen 0.4 routes every GroupChat turn through a typed pub-sub bus, with the manager publishing SpeakerRequest events that trigger the chosen agent's reply.
Think of a TV game show. Every contestant has a buzzer that only lights up when the host calls their name. Contestants do not shout over each other and do not decide who goes next. They wait for the host. When the host points at contestant B, only B's buzzer lights, B answers into the shared microphone, and everyone in the studio hears it. The host then decides who goes after B based on what was just said. AutoGen 0.4 builds that exact setup in software. The bus is the studio sound system, the manager is the host, and the SpeakerRequest is the host pointing at the next person.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
AutoGen 0.4 is not a minor version bump on 0.2. It is a full architectural rewrite that replaces the original conversational-Python-object model with a typed actor runtime sitting on a message bus. SelectorGroupChat is the smallest example that exercises every interesting part of that runtime, which is why interviewers reach for it.
The question being probed is whether you understand that there is no direct call from manager to worker. Everything is a publish followed by a subscriber waking up, and that single property is what lets the same code run as a unit test in one process and as a fleet of containers behind a gRPC gateway in production.
Mental model: the manager and the agents do not know each other's identities. They know topic names. The bus does the routing.
The actor runtime and why it exists
What changed from 0.2
In AutoGen 0.2, a GroupChat was a Python object that held references to agent objects and called their generate_reply method directly. That made small demos easy to write and made distribution, replay, and isolation hard. The 0.4 rewrite moved every agent behind a typed inbox served by an AgentRuntime. You no longer call an agent; you publish a message and the runtime delivers it to whoever is subscribed.
Two runtimes, one programming model
The SingleThreadedAgentRuntime runs the whole bus in one Python process and is what you use for tests and notebooks. The distributed runtime backs the bus with gRPC workers and a host gateway, so agents can live in different processes or different machines. Crucially, the agent code is identical between the two. The actor model is the abstraction that makes that possible: agents only see typed messages on their inbox, not pointers to other agents.
Why this matters for GroupChat
It means GroupChat is not a hard-coded conversation loop. It is a recipe for which topics to wire and which messages to publish. SelectorGroupChat, RoundRobinGroupChat, and Magentic-One are all just different recipes over the same runtime.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
from autogen_agentchat.teams import SelectorGroupChat
from autogen_agentchat.conditions import MaxMessageTermination
from autogen_agentchat.agents import AssistantAgent
from autogen_ext.models.openai import OpenAIChatCompletionClient
model = OpenAIChatCompletionClient(model="gpt-4o")
planner = AssistantAgent("planner", model_client=model,
system_message="You plan steps and hand off to coder or critic.")
coder = AssistantAgent("coder", model_client=model,
system_message="You write code only.")
critic = AssistantAgent("critic", model_client=model,
system_message="You review code and say APPROVE or request fixes.")
team = SelectorGroupChat(
participants=[planner, coder, critic],
model_client=model,
termination_condition=MaxMessageTermination(12),
)
# Every reply is a TextMessage on the bus.
# Every speaker switch is a SpeakerRequest from the manager.Real products, models, and research that use this idea.
- Microsoft AutoGen Studio runs the same SelectorGroupChat over the in-process runtime for prototyping and the distributed runtime for hosted demos.
- Magentic-One (Microsoft Research, 2024) is a SelectorGroupChat composition with WebSurfer, FileSurfer, Coder, and ComputerTerminal agents on the same bus.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does SelectorGroupChat avoid picking the same speaker forever?
Tune the selector prompt to require variety, set allowed_repeated_speaker=False in the selector config, add a MaxMessageTermination as a hard cap, and consider a custom selector_func that enforces a minimum gap between repeats.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Treating the manager as a Python function call instead of a subscriber on the bus, which hides why the same code can run distributed without changes.
60 second bullets to scan on the way to the call.
What an actor is in the AutoGen 0.4 runtime and how it differs from a 0.2 agent
Which topics SelectorGroupChat uses (chat TextMessage plus per-agent SpeakerRequest)
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.