The concrete moment in Katie Paul's Reuters account is an engineer reassigned to write software puzzles for a model to train on, calling the job rote in an internal post [23]. Around the same time Meta mandated tracking software on US employees' devices to capture keystrokes and mouse clicks, so agents could learn to use a computer the way a person does, and paused that programme in June [24]. Staff answered executive posts with pictures of elephants [26]. Internal sentiment in the half-year Pulse survey went from 74% favourable to 55% [27], a drop of 19 points [3].
Andrew Bosworth's two counters are the useful part, because they were published by the company running the experiment. Dividing the growth rates, 220 over 36, internal activity expanded about six times faster than the slice of it a user could open [1]. Firefighting hours grew nearly twice as fast as the incident count, 70 against 40 [2], which is what it looks like when failures get harder rather than merely more numerous. An April post had already described unchecked agents taking "large-scale, disruptive actions that humans are unlikely to execute" [17], infrastructure teams flagged reliability warning signs in March [20], and in June attackers used Meta's new AI customer support bot to reach high-profile Instagram accounts, among them the dormant Obama White House page [21]. Meta declined to comment on the internal data [22].
What the plan told itself is visible in its design: AI would take over daily work performed by thousands, supervised by smaller cadres of what one planning document called "talent-dense" staff [4]. What happened is that the supervision did not stay inside the org chart. It arrived as incident response, owned by infrastructure and security people, while the teams being sized down sat elsewhere. One human-resources executive had projected a culling as big as or bigger than the roughly 25% Meta cut three years ago [6]; the wave that actually shipped was about 10% [8], under half that scale [4].
The provenance is worth keeping in view. Executives had studied how AI startups organise, and chief data officer Alex Schultz and head of product Naomi Gleit both visited Asia [13]; Gleit told Reuters in June that time in Meta's Singapore office had "inspired some of the teams in California and New York", and that many of the ideas were bottom-up [14]. Meta's own framing, that it "didn't move forward with every scenario from the exercise, and it was never assumed we would" [12], is the standard defence of a planning document, and it holds up better than usual here, because Reuters could not establish what changed Zuckerberg's mind that night [9].
A test that travels to smaller companies: put two counters on the same slide, work produced and work that reached a user, and one cost line for hours spent cleaning up after the agents. If the second counter does not move with the first, the programme is buying volume. If the cost line sits with a different team than the one being reduced, the saving is a transfer between budgets rather than a saving.
