Review is the step that looks like pure overhead when throughput is the number on the wall. It adds latency to every change, it ships nothing by itself, and its output is a question rather than an artifact. It is also the only step that asks whether a capability should exist at all, which is a question no generated test will pose, because a test asserts behaviour somebody already imagined. Engineers at Meta and Instagram told the Pragmatic Engineer that the offending change was AI-generated and AI-reviewed [5], and a generator and a checker drawn from the same model share the same imagination.
The reassignments described in the same reporting are the more legible trade. Integrity staff at WhatsApp were moved onto enforced data-labeling work [8], and forced reassignment off Integrity teams onto AI labeling and related duties is one of the causes the engineers named [5]. The headcount figures around them are steep: a 95% reduction on Developer Documentation and Support [7] leaves that team at a twentieth of its former size, and the 44% cut to Instagram's design team [6] leaves it at 56% [14]. Neither of those is a security team, but both are places where somebody writes down what a system is supposed to do.
A skeptic would say one humiliating bug does not indict a strategy, since platforms of this size ship regressions constantly and this one was closed. The answer is in the detection path rather than the defect. According to the Pragmatic Engineer, Instagram's dedicated Integrity team appears to have discovered the problem through news coverage [3]. An organisation that catches its own capability errors pays for them in review hours. An organisation that does not pays on a schedule set by whoever publishes first.
The incident is better evidenced than its explanation. The takeover of high-profile accounts, including that of former US president Barack Obama, is reported as fact [2]; the causal chain runs through unnamed engineers speaking to a single newsletter, and no Meta account of the cause appears in that record [15]. Read the first as established and the second as testimony from people with a view. The more portable finding is the newsletter's trend read: individual productivity up, team productivity flat [10]. That is the shape you would expect if the per-engineer gains are real and are being absorbed downstream by rework and by the reviews that did not happen earlier.
It is worth keeping the clock straight. The Pragmatic Engineer dates the change to around November and a more capable generation of agents including Opus 4.5 and GPT-5.4 [11], which makes this a six-month problem in tooling, not a decade-long verdict on how software gets written. What a leader setting next quarter's targets can take from it is a trade-off stated plainly: agent-assisted output per engineer shows up in a dashboard within weeks, while the cost of an unreviewed capability arrives whenever an outsider finds it. Cut the gate to make the first number move, and the second number you learn about is time to detection, reported by someone who does not work for you.
