Engineering counts merged pull requests and nothing for the hours spent watching the agent

No engineering dashboard has a column for the hours in Quentin Rousseau's account, and that is the operative problem. The Rootly CTO describes watching an agent as passive enough to feel like rest, active enough to keep you hooked ^[6]. The pull request that arrives at the end of the night is countable. The night is not.

The supply side of that loop is at least documented. Some 45 percent of Stack Overflow's respondents report frustration with AI answers that are almost right but not quite, output that reads as convincing and debugs as expensive ^[8]. ZDNet's write-up gives the residue a name, verification debt: the code lands fast, and the developer still owns the question of whether it is correct, secure, maintainable and appropriate to that particular codebase ^[9]. That work produces no artifact.

Then the amplifier. Where an employer treats agents as multiplied developer capacity, the expectation becomes more features shipped, more tickets closed and more reviews performed in the same number of hours, which can consume the time saved on individual coding tasks ^[11] and reappear as larger pull requests, more generated changes to inspect and more dependencies to validate ^[12]. Every item in that second list is countable, and all of it counts in the direction that flatters the tool. The 39 percent of developers who say AI has made it harder to switch off from work ^[14] are reporting the other side of the same ledger, and nothing is pointed at it.

The evidence deserves a discount. Coddy Tech sells programming training, and its sample is 305 self-reporting developers ^[1], which means a single percentage point is worth about three people ^[13]; the precise figures should not travel into a planning document. The direction survives the discount, because a separate instrument shows the same divergence: in the Stack Overflow survey, trust in AI accuracy fell 11 points in a year and favorability fell 12, while adoption kept climbing ^[16].

A manager who wants a number here has to build it, because standard telemetry does not emit one. The nearest honest candidate is a ratio: hours spent supervising and repairing generated work against hours spent authoring, self-reported, per person, per week, sitting next to the count of merged changes. ZDNet's own conclusion is that agentic programming is becoming as much a life-work balance question as a tooling one, and that teams using agents to remove routine toil may see real benefit ^[15]. That distinction is only auditable if someone measures the supervision. Until then, the productivity case rests on the half of the ledger that happens to be instrumented.