An agent kept retrying a failed step for hours and nobody knew. What we now require before anything touches live data.
A thought to take with youIf you cannot see what an automated system did and why, you cannot trust it with real work.
What we expected
In an early prototype, we relied on standard error handling: if a connected system returned an error, the step would fail and someone would be told.
What happened
The connected system reported success while hiding a failure inside its response. The agent kept retrying for hours. We only noticed when the day’s numbers did not reconcile.
What we changed
Every step now leaves a readable record of what it did and what came back. Agents have limits on retries, time and spend, and they stop and alert a person when a limit is reached. This is in place before any agent touches live data.