Agent loops can fail without errors

A study of 147 incidents across 81 production runs found one agent loop repeat the same tool call 54 times in 11 minutes, even though every call returned success. In the Agent Mesh failure study, ordinary retries, timeouts, and error-rate breakers miss non-idempotent delegation; the costliest enforcement failure consumed 107 agent turns and produced zero accepted writes. If an agent can mutate systems, track business effects and forward progress—not only HTTP failures—before adding more retries, a boundary also documented in the archive’s agent control-plane analysis.