The 90-second diff still costs 20 minutes of human review

Put the post's two figures on the same line. If a working day is eight hours, then 480 minutes of writing drew about 20 minutes of review, and review was four percent of the elapsed time on a change ^[20]. Ninety seconds of generation against the same twenty-minute review makes review about 93 percent of it ^[20], and review then takes roughly thirteen times as long as production ^[22]. Both figures are illustrations in the post, not logged measurements.

"Generation got cheap. Verification did not," Debashish Ghosal wrote on dev.to ^[2]. The complaint he says he keeps hearing from engineers is "I spend more time reviewing AI code than I ever spent writing code myself" ^[3].

The design detail I would copy is where the authorization sits. His gate has four states, allow, audit, escalate and deny, and it runs outside the model, so no prompt can override a deny ^[7]. Binary allow and deny cannot tell a sandbox delete from a production delete ^[8]. A team holding only those two states picks between over-privileged agents and approval fatigue ^[8].

Gate 2 is the one with a number attached. Structural validation runs before human review: every precondition maps to an earlier task, and every step with a large blast radius has a reachable rollback. Ghosal says this is deterministic, costs nothing, and eliminates roughly half the defects before a person looks ^[10]. For that half to hold on your repository, your defect population would have to be dominated by ordering and rollback gaps, because ordering and rollback are what these checks test. He is direct about the limit: the structural checks are weak on logic that is subtly wrong in a way nobody anticipated ^[18]. The post gives the sizes of his field tests, 83 real agents across 10 frameworks and 170 goals, without the pass and fail counts ^[7][9].

The third gate is a check on the checks. Ghosal compares a deterministic gate that silently stops firing to a Kubernetes ReplicaSet that kept one pod Running while four releases never actually ran, with every health check passing ^[12]. His fix is a canary that asserts each gate still fires, so that a dropping blocker count is not the first sign a guard died ^[13].

A deny needs a matching path out. An agent that refuses 96 of 97 risky plans, about 99 percent ^[21], is doing its job only if the escalate route reaches a human with enough context to decide. Without that route, the deny just stops the work ^[17].

Review attention does not scale linearly with volume, and the post describes the way it degrades: at high volume people pattern-match on plausibility instead of reasoning about correctness, and the cleaner the diff, the less likely it gets interrogated ^[14]. In my view that is the strongest claim in the piece, because it predicts which changes get waved through. Ghosal's conclusion is that "You don't fix a verification bottleneck by adding humans at the same step" ^[16].