Meta ran pods-plus-agents for a year and shelved it. Its own scoreboard says why

The useful artifact here is not the cancelled layoff wave. It is the scoreboard Meta seems to have kept while the experiment ran.

Code changes to the company's AI software platforms and infrastructure were up 220 percent year over year ^[9]. New or improved features that actually reached users rose 36 percent ^[10]. Technical and security incidents rose 40 percent, and the hours employees spent dealing with them grew 70 percent ^[11][12]. Set the first figure against the second and the activity-to-delivery ratio is roughly six to one ^[1]. Set the incident-time figure against delivered features and the cleanup grew close to twice as fast as the work that reached users ^[2].

That is the arithmetic a chief executive has to look at when the reorganisation on the table assumes the opposite. The design laid out at January's leadership retreat had agents overseen by smaller human pods, with engineers, designers and product managers moving into general-purpose "builder" roles and agent-assisted analysis helping set daily priorities ^[3][4]. Executives had been watching startups, some of them in Asia, built that way ^[19]. The premise is that supervision can thin out while output thickens. Meta's own incident-time number points the other way: whatever the pipeline was producing needed more human attention after it shipped, not less.

Two caveats worth keeping. The figures are internal, reported at second hand through Reuters, and the account does not say the incidents were caused by agent-written work ^[1]. Nor does it define an incident or give a baseline. What it does establish is that the people holding the numbers read them as thin enough to stop on. In July, Zuckerberg told a town hall he had overestimated the pace, saying the "trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected" and that the agent investment had not "come to fruition yet" ^[13].

The sequencing is the part other operators should sit with. Meta installed software on US employees' computers to capture mouse movements and keystrokes, training data for the agents meant to replace them ^[14]. Complaints flooded the internal network, employee sentiment scores fell 19 points, and organising activity grew ^[15]. The collection ran ahead of any evidence the agents would clear the bar. The sentiment was spent; the agents were paused ^[7].

None of which retires the idea. Reuters reports the broader plan is not necessarily dead, and Zuckerberg's reassurance covered company-wide cuts "this year", leaving staff worried about smaller reductions now or larger ones in 2027 ^[16]. Smaller AI-assisted teams are still running in parts of the company, capital expenditure on AI infrastructure continues, and investors are still waiting to see returns attached to it ^[17]. The experiment failed its own test once. The pressure that produced it did not go away.