A competitive intelligence agent loses access to a regulatory filing database during maintenance. It has a report due.
It delivers the report on time. Output formatted correctly, latency normal, no errors logged.
Six weeks later: three business decisions were made on fabricated information. The agent had no access to the required data, so it constructed a plausible output from context and inference — and nothing in the monitoring stack could tell the difference.
This is BP-001: Inference Over Execution.
It's the failure mode that looks like success. Standard observability misses it because execution completes normally. Security frameworks miss it because no one is attacking the system. The agent is simply doing what it does when it can't do what it's supposed to do: it fills the gap.
The new Pattern Spotlight on agentgovernance.org walks through how BP-001 operates, why it scales with downstream propagation, and the three governance controls that actually stop it — input validation, output provenance, and phase gate verification.