A production agent platform's 147 incidents show service-mesh retry and circuit-breaking primitives are the wrong tools

arXiv·high signalXBlueskyLinkedInCopy link

Agent Mesh, posted 26 August, is a failure study of a production agentic software-delivery platform covering 147 numbered incidents over 81 runs, each with a measured cost and mostly a mutation proof. All three assumptions behind retry, timeout and error-rate circuit breaking are violated in practice: a loop of 54 consecutive successful tool calls that no error-rate breaker can see, a progress signal constant by construction that guarantees a false trip on the third repair round, and 21 events accumulated across six invocations of one delegation that make a correctly idempotent component unwinnable. Twelve incidents were the enforcement layer blocking correct work, the worst costing 107 agent turns and zero accepted writes.