Meta built a plan to shrink parts of its workforce by as much as 60 percent and replace the lost output with AI agents, then reversed course hours before executing the second phase. Internal documents describe the effort, codenamed Project OT, as a restructuring that would leave small “talent-dense” teams supervising fleets of virtual workers. Reuters reported the reversal on August 26, citing those documents and internal communications.
The stakes go beyond one company’s headcount math. Meta is one of the frontier AI labs testing agentic labor substitution on its own staff, and the outcome undercuts a specific claim: that agents are ready to absorb the operational load of “talent-dense” teams today, not eventually.
Timing sharpens the story. Just hours before the first layoff round went out on May 19, Zuckerberg called off the plan’s second phase, which had been penciled in for November. Reuters attributes the reversal to three forces acting together: the agents failed to produce the output gains Meta had projected internally, investors pushed back on the scale of AI spending, and staff pushed back hard against the plan itself.
Zuckerberg had already flagged part of this in July, telling employees that agent capability was advancing more slowly than he had anticipated. Pair that admission with the August reporting and a pattern emerges: Meta’s internal targets for agentic automation outran what its own systems could actually deliver, and the shortfall was wide enough to force a public reversal rather than a quiet slip in the schedule.
A second, separate failure runs alongside the technical one. Employees concluded that software logging their clicks and keystrokes for productivity tracking was doubling as training data for the systems meant to replace them, Reuters reported, and answered by posting objections across Meta’s internal channels in large numbers. Sentiment inside the company, measured over the same stretch, dropped from 74 percent to 55 percent.
That sentiment drop is the more durable signal of the two. A failed productivity target can be revised with a new roadmap and a longer timeline. A workforce that concludes its own monitoring data is being repurposed to eliminate its jobs does not return to its prior trust level once the specific project is paused, because the underlying incentive (tracking software feeding a replacement system) does not disappear with the announcement.
Meta has not disclosed whether Project OT itself is cancelled or merely deferred, nor has it said whether the monitoring software identified by employees will continue to run. Reuters did not report a revised timeline for any future wave. That absence is itself informative: a company confident in a near-term fix for the agent shortfall would typically attach a revised date, and none has surfaced.
Meta set an explicit headcount-reduction target tied to agent performance, missed the internal benchmark, and reversed under combined technical and workforce pressure. Any operator running a similar substitution plan, tying a restructuring timeline to an agent capability curve rather than a measured production benchmark, should treat this as evidence that the gap between agent demo performance and agent production performance can still be wide enough to blow up a public commitment. Companies disclosing monitoring tools to staff should also expect that employees will infer, correctly or not, what the data is for.
Reuters, in reporting reviewed by The Decoder, first detailed Project OT’s collapse on August 26, 2026.
