How is the termination condition expressed in an AutoGen GroupChat?

**AutoGen GroupChat** has two termination layers that work together. First, `is_termination_msg`. A small predicate function you attach to each agent (and to the GroupChatManager). After every message, AutoGen calls this predicate on the last message; if it returns True, the agent stops sending replies. The conventional pattern is to instruct the assistant in its system prompt to write the literal word 'TERMINATE' on its final message, and to wire the predicate as a check for that sentinel in the message content. Second, `max_round`. A counter on the GroupChat that caps the total number of turns. Even if no message ever matches the predicate, the chat ends after `max_round` turns. It is the safety net for runaway loops. The two layers are complementary. The predicate produces clean, intentional endings. The round counter prevents pathological cases. There is no `register_termination_hook` API; the manager LLM is not by default asked to vote on termination at each step. That is something you can build, not the default behaviour.

Termination in **AutoGen GroupChat** has two layers. - **`is_termination_msg`**. A predicate `Callable[[dict], bool]` attached to each agent and to the GroupChatManager. After every message, AutoGen calls it on the last message; True stops the loop. The conventional pattern: instruct the assistant to write `TERMINATE` on its final message and check `lambda m: "TERMINATE" in (m.get("content") or "")`. You can substitute a stricter check (exact match, JSON parse) or even an LLM-based predicate. - **`max_round`**. A hard ceiling on the GroupChat. After this many turns the loop ends regardless of predicates. Both layers matter. The predicate is the intended exit and produces clean endings. The ceiling is the safety net. A misbehaving model that never emits the sentinel still terminates rather than running forever and burning tokens. The wrong answers in the question are wrong for distinct reasons. `max_round` alone leaves no clean way to stop early. Asking the manager LLM to vote at each step is something you can build but is not the default. There is no `register_termination_hook` API. That name does not exist in the framework.