Agentic debugging: does the operator choose a control before the agent reads code?

Research brief, October 21. Question: does a recent, firsthand production record show that a human can steer agents through an unfamiliar, maintained codebase without taking a model’s first plausible diagnosis on trust? A good answer names each observed failure, discriminating control, rejected hypothesis, validated repair, and what was actually deployed or checked. We already have a released fix without reporter retest, green tests with wrong browser fixtures, device acceptance and player-triggered redesign. Subquestions: (1) What do three bug hunts demonstrate beyond the first latency case? (2) What evidence distinguishes a causal fix from a plausible patch? (3) What human attention and deployment constraints make this hard to generalize? (4) Does more raw observability really improve autonomy, or transfer a new curation and privacy burden?

Source and status. In his October 21 fragment, Martin Fowler points to Cindy Sridharan’s long first-person essay “How to effectively debug when you no longer understand the code.” His summary says three bug hunts motivated comparing a failing case to a working control, separating raw observations from hypotheses and testing each candidate against all facts. He reports that the agent did not choose the initial control in one latency investigation until the human suggested it. This is Fowler’s reading of Sridharan, not yet our reading of her original. The original link could not be verified from the accessible text; do not treat any guessed Medium URL or publication date as verified. Next: find the author's canonical link, read all three cases, check whether outcomes reached deployed behavior and record the limits. The 64-minute essay is weekend-only, after prior hearted originals if capacity. Do not post Fowler’s summary in place of the author’s investigation.