Autogen Termination Conditions

Team uses TextMentionTermination('DONE'). Agent emits 'we are done with the task' (lowercase 'done'). Team does not stop.

TextMentionTermination matches the configured substring as-is — semantics are case-sensitive substring containment. Choose unambiguous sentinels (e.g. '<<TERMINATE>>') that won't appear in normal prose, and prime the agent's system_message to emit the exact token at task completion.

Team uses TokenUsageTermination(max_total_token=20000). One model_client doesn't report usage and the counter stays at 0.

TokenUsageTermination relies on the model_client surfacing per-call usage. Verify your model_client populates the usage field on every response (some custom clients may not). Without usage signal, the termination never fires — pair with MaxMessageTermination as a backstop.

Team has MaxMessageTermination(max_messages=10). Operator expects 10 assistant messages but the team stops after ~7.

MaxMessageTermination counts all messages in the team's transcript — tool messages, user proxy turns, system messages count too. Set the cap based on total transcript length, not assistant turn count. Inspect the transcript shape (text + tool + tool-result) when calibrating the cap.