The agentic harness engineer’s day is four loops overlapped, not one. • Instrument: Add or refine trace fields so a regression can be found and reconstructed. • Cluster: Group production failures into named regression buckets, not one bug list. • Ship: Change prompts, tools, retrieval, or halt conditions against a named cluster. • Verify: Grade the change through the eval suite before it ships to real users. Four loops running at once, on different cadences. Instrumentation happens once a week, clustering daily, shipping several times a day, and verification is continuous. The failure mode is running all four as one loop, triggered only when a customer complains. That is the reactive mode, and it is how an agent slowly fails/drift in production.