The Agent That Never Said It Was Stuck

An autonomous inventory-reconciliation agent, deployed to resolve discrepancies between a WMS and physical counts, ran for eleven hours without producing output.

The postmortem found the agent had entered a retry loop on a single malformed record at hour one, and every subsequent 'status check' the agent performed reported the same generic 'in progress' signal — because the agent's internal state genuinely was 'in progress,' just on the same step, forever.

The failure wasn't a bug in the retry logic. It was the absence of a distinct signal for stuck versus working — the agent had no way to represent, even to itself, that time was passing without progress.

We now treat 'progress observability' as a first-class reliability requirement for any agentic system we evaluate, not an optional logging nicety.