AI makes experts slower: why approval must be an architecture, not a chore

The METR study measured experienced developers as 19% slower with AI, while they believed they were 20% faster. Review and verification consumed the gain. The problem is not review: it is where we put it.

The METR study measured experienced developers as 19% slower with AI on their own code, while they believed they were 20% faster (newsletter.getdx.com/p/metr-study-on-how-ai-affects-developer-productivity). The gap comes from review and verification consuming the apparent gain. Our thesis: the problem is not review, but where it is placed.

What the METR study actually measures

The contrast is clear: ‘developer self-reports after using AI are overoptimistic to the point of being on the wrong side of speedup/slowdown’, says the study's lead author in the Hacker News thread sharing it (news.ycombinator.com/item?id=44522772). Rereading AI-produced code line by line to check it does what you think you requested costs more than writing it yourself in many cases. This is not a surprise; it is review at the wrong level.

Line-by-line review does not fit into a normal working day

A non-developer cannot review code line by line anyway: it simply is not an available option, however willing they are. But even an experienced developer loses time doing it, according to METR, despite having the skills and vocabulary. The review that matters concerns the intention before it becomes code: the brief, specification and story breakdown. A short document can be judged in minutes; code takes hours to review, still without any guarantee of seeing everything, especially when hundreds of lines change at once in files you have never opened yourself.

What Maestro moves

Three mechanisms concentrate verification where it is effective rather than where it is customary. First, the ‘Looks good to me’ gate: you approve short documents at the right milestones, never code. Next, tests written before code: technical verification runs automatically without you having to read it. Finally, review by a dedicated agent, bounded in scope rather than promised to be infinite, never replaying tests or verification already completed so as not to duplicate lengthy work done elsewhere in the chain.

Proof through the harness

This shift in approval is more than a promise: it is what our own tester agent checks, playing the entire journey in place of a human user before every update, with several temperaments included. Approval exists, but it concerns the right object at the right moment, not every line of every file written by an agent.

What it means for you

You do not escape verification by directing an agent team; you change its nature. A report cited by SQ Magazine notes that the best gains from AI agents concern tasks with clear success criteria and recoverable errors (sqmagazine.co.uk/ai-agent-autonomy-statistics): exactly what a well-placed gate produces, rather than general review. A clear success criterion written in a short specification does more for product reliability than hours of review without a rubric or method.