Meta's agent plan produced 220 percent more code changes and 36 percent more shipped work

The load-bearing figure is the ratio between those two code numbers. Internal churn grew roughly six times as fast as work that reached users ^[25]. That is what it looks like when agent output is measured at the commit, because a commit is not a product.

The repair side compounds it. Growth in response time outran growth in incident count by enough that the average incident consumed about 21 percent more staff hours than a year earlier ^[20]. So the failures were both more frequent and individually harder to close, and Meta's infrastructure teams tied reliability problems to the surge in AI-generated code ^[15]. That work lands on engineers reading code they did not write, at volume, under an outage.

Set that against the one case Meta can show working. In April its engineering organization said an internal system compressed roughly 10 hours of performance-investigation work into about 30 minutes and produced code changes for human review ^[21]. Call it a factor of twenty ^[26], on a bounded task with a checkable answer and a human at the gate. Project OT tried to extend that model to management, product development and other less structured work before Meta had evidence the technology could carry it ^[22].

Note what Meta actually contests. It confirmed Project OT and the two-wave structure, confirmed that the most aggressive scenarios contemplated shrinking some teams by up to 60 percent, and said the figure never applied to the whole workforce, that several major units were excluded, and that leaders dropped the second wave before deciding its size ^[7]. Every one of those points is about scope. None of them is about the productivity data.

The plan also carried a labor bill nobody costed. Meta reassigned some engineers to produce software-engineering tasks used as training data for its coding models ^[8], and required tracking software on US employees' devices to record keystrokes and mouse activity so agents could learn how people use computers ^[9]. Employees concluded they were training their replacements and filled internal channels with criticism, according to Reuters as summarised by the-decoder ^[10]. The same account lists investor complaints about the size of the AI budget among the triggers for the reversal ^[11], and reports that in July Zuckerberg said the agent technology had not sped up as fast as he expected ^[17].

What did happen was ordinary. Meta went ahead on May 20 with a reduction affecting about 10 percent of its workforce ^[6], and between layoffs and transfers some engineering units were down as much as 30 percent by the end of May ^[23]. The second-quarter filing counted 75,472 employees as of June 30 and said that total still included about 8,000 affected people ^[24], roughly 11 percent of the headcount on the books ^[27]. Reuters built its account from internal documents, recordings and interviews with more than 20 people ^[12]. The cuts were delivered on schedule; the substitution they were supposed to make permanent was not.