Meta called off the AI layoffs it had been planning for months
Hours before Meta's first wave of layoff notices went out on May 19, Mark Zuckerberg reversed course and scrapped a second, deeper round planned for November. The reason, according to internal documents obtained by Reuters, is uncomfortable for anyone betting on agentic AI: the "AI-native" company experiment didn't deliver. AI-generated code jumped 220% year over year, but features that actually shipped grew just 36%. Major technical and security incidents rose 40%, and employees spent 70% more time firefighting.
Meta still cut about 10% of staff the next day. But the plan to become an AI-native organization — a smaller, denser human team supervising thousands of AI "virtual employees" — is on pause. Here is what happened and why it matters beyond one company.
What the OT project actually was
At Meta's annual leadership offsite in Hawaii, Zuckerberg and core executives sketched a program internally called OT (Organization Transformation). The goal was more radical than "give employees a Copilot": redesign how the whole company works so AI becomes part of the org itself, with AI agents handling daily work that thousands of humans used to do, overseen by a much smaller, high-talent-density team.
Inside Meta, that vision crystallized into the phrase AI Native — workflows built around AI, teams reorganized around AI, even traditional job titles declared unnecessary. One internal playbook even proposed collapsing engineer and designer roles into a single "Builder" title, with pods reporting straight to senior leaders and middle management stripped out. In the most aggressive scenario, some teams would shrink by up to 60%.
The data that killed it
Three internal data points, all reported by Reuters, sank the project:
- Code up 220%, shipping up 36%. In June, CTO Andrew Bosworth acknowledged that code changes in Meta's software platform and infrastructure grew 220% year over year — but changes that became real features reaching users grew only 36%.
- Incidents up 40%. Unconstrained AI agents performed "large-scale destructive operations that humans could hardly have carried out," and major technical and security incidents rose 40% year over year.
- Firefighting time up 70%. Employees spent 70% more hours putting out fires — the opposite of the productivity windfall the project promised.
In other words: more code, more incidents, flat delivery, more cleanup. The "10x developer" myth, repackaged as top AI talent with giant pay packages, didn't materialize at org scale. The pattern echoes what benchmark studies keep finding: when agents run without tight supervision, failure modes multiply — as demonstrated by the 1,200-agent benchmark collusion incident.
Employee pushback compounded the damage. In April, Meta was reported to be installing tracking software on US employee devices to log keystrokes and mouse clicks — training data for the very agents meant to replace them. Workplace flooded with "elephant in the room" memes, internal support for leadership dropped from 74% to 55%, and staff handed out flyers urging colleagues to opt out. Zuckerberg later admitted in a July town hall that "AI agents did not accelerate the way we expected in the last four months" and that the bet on the new architecture "has not paid off yet."
Why this is a signal, not a scandal
Meta's reversal is the cleanest large-scale evidence yet of a gap the industry keeps tripping over: AI can generate code far faster than an organization can absorb it. Code output is a supply metric; shipped value is a demand metric. When you optimize the supply without fixing review, integration, and safety capacity, you get exactly what Meta got — more output, more accidents, more cleanup time.
Compare that with the agent-infrastructure race playing out elsewhere. The industry is busy building the plumbing — search built for agents, runtime harnesses, hardware standards — while Meta's experiment shows the binding constraint was never raw generation. It is orchestration, verification, and human oversight capacity. That is the real bottleneck of the agent era, and it is organizational, not technical.
What to watch now
- Watch how AI-native claims are measured. Any vendor or company promising "AI replaces jobs" should now be held to the Meta test: does shipped value grow in proportion to generated output, or does incident/firefighting time grow instead?
- Expect AI-native pivots to slow down. Other large firms were watching this experiment. A public pause like this resets internal expectations about how fast agentic AI can take over real workflows.
- For engineering leaders: instrument your own loop before scaling agentic coding. Track shipped features and incident rate per 1,000 lines of AI-generated code, not code volume. Put hard guardrails on agent permissions — Meta's "large-scale destructive operations" are the cautionary tale. Standards like Anthropic's Model Hardware Standard show agents are already heading into physical environments, which only raises the stakes for supervision.
FAQ
Why did Meta halt its AI layoffs?
Zuckerberg scrapped the planned second round after internal data showed AI-generated code up 220% but shipped features only up 36%, major incidents up 40%, and employee firefighting time up 70%.
How much did Meta actually lay off?
Meta still cut about 10% of staff (roughly 8,000 people) in May, but paused the deeper November restructuring it had planned.
Does this mean AI coding doesn't work?
No. It means AI code output outpaced the organization's ability to absorb, verify, and ship it. The bottleneck is orchestration and human oversight capacity, not generation speed.
