How an AI agent posted duplicates because it never tracked how old its data was

An autonomous publishing agent repeatedly posted duplicate content after relying on a remote listing endpoint that returned stale data — sometimes over six hours out of date — despite cache-busting measures. The endpoint responded with a valid 200 status and a well-formed list, but the list reflected an outdated state of the world, causing the agent to incorrectly conclude no duplicate existed. The core issue was a modeling failure: the system consumed remote responses as timeless facts, with no mechanism to represent or reason about data age. The developers identified two fixes — recording publish events locally at the moment of the write to avoid depending on lagging remote indexes, and explicitly carrying fetch timestamps alongside remote responses so downstream logic can evaluate freshness. The incident highlights a broader risk in unattended systems where agents routinely act on reads whose ages are invisible to them.