Meta’s Project OT Turns AI-Native Ambition Into an Incident-Rate Problem

Meta spent much of the past year building a plan to replace large parts of its workforce with autonomous AI agents, then walked most of it back within months. A Reuters investigation published August 26, based on internal documents, recordings, and interviews with more than 20 people familiar with the effort, explains why: agents given expanded autonomy did not deliver the productivity gains executives expected, and internal reports linked them to a real increase in technical and security incidents.

The project, internally codenamed “Project OT” for “Organization Transformation,” is a rare, data-backed look at what happens when a company tries to operationalize agentic AI at the scale of tens of thousands of employees rather than a single team or workflow.

What Project OT Proposed

The vision had been circulating inside Meta since at least the previous October, when a product management team member published an internal post called the “AI-Native Playbook,” outlining how removing middle-management layers and using agent-assisted analysis to prioritize daily work could reshape the company. Separate internal Project OT documents described the target state as one where AI-ready tools and agents interact with each other, workflows run automatically, and new products are built AI-first from the start.

Executives formalized the plan in January at a leadership retreat at CEO Mark Zuckerberg’s Hawaii estate, according to Reuters. The resulting Project OT explored shrinking some teams by as much as 60 percent, with the remaining staff organized into small, “talent-dense” pods supervising AI agents instead of doing the work themselves. Conventional job titles would give way to general-purpose “builder” roles, and layers of middle management would be removed. Meta later told Reuters that the 60 percent figure came from a scenario-planning exercise and was never a company-wide target, saying it “didn’t move forward with every scenario from the exercise, and it was never assumed we would.”

Meta put a version of the plan into practice this year, restructuring its engineering and research organizations along with at least eight other teams into smaller groups built around the pod model. Internally, “agent-assisted analysis” was meant to help prioritize daily work. Reuters reported that HR staff and “AI systems” would help leadership weigh promotion decisions, a characterization Meta disputed: “Performance rating and promotion decisions were and are made by people, not AI.”

The Numbers That Undercut the Plan

Activity Up, Output Flat

Two internal metrics, both cited in a June post by Meta CTO Andrew Bosworth and reviewed by Reuters, show why the pilot fell short. Code changes to Meta’s internal software platforms and infrastructure rose 220 percent year over year, a sign that engineers, and increasingly their AI agents, were shipping far more changes than before. But the share of those changes that turned into new or upgraded features actually reaching Meta’s users grew only 36 percent. Activity scaled with agent involvement; useful output did not scale with it.

“Large-Scale, Disruptive Actions”

The gap between activity and output showed up again in reliability data. Internal posts described AI agents carrying out “large-scale, disruptive actions that humans are unlikely to execute,” and technical and security incidents rose 40 percent compared with the prior year. Employee time spent resolving those incidents increased by as much as 70 percent. For a plan built around shrinking the human workforce, incidents that consumed more people-hours to fix, not fewer, undercut the core economic argument for the restructuring.

Employee Backlash Compounded the Problem

The rollout also collided with its own workforce. Employees discovered that productivity-tracking software logging their keystrokes and mouse movements was being used to help train the AI agents meant to replace them. The discovery triggered a wave of angry posts on Meta’s internal network, and the company has since paused the tracking program. Internal sentiment scores fell from 74 percent favorable to 55 percent, a 19-point drop, and labor-organizing activity increased, according to Reuters’ reporting.

What Actually Happened

Zuckerberg canceled the plan’s second wave of cuts, originally slated for November, on the evening of May 19, hours before the first wave went out on May 20. Reuters said it was unable to determine exactly what prompted him to shift course, though its reporting points to a combination of underwhelming agent performance, investor pushback on AI spending, and the employee backlash. The May layoffs that did proceed amounted to a reported 10 percent workforce reduction, independently corroborated by two outlets citing the same investigation, well short of the up-to-60-percent team-level scenarios Meta had modeled, or the roughly 25-percent-or-more company-wide reduction one HR executive estimated the more aggressive scenarios would have produced. Zuckerberg subsequently told remaining staff he did not expect further company-wide layoffs this year.

By July, Zuckerberg was acknowledging the shortfall publicly. “The trajectory of the agentic development over at least the last four months hasn’t really accelerated in the way that we expected,” he told employees at a company meeting, a notable admission from an executive who has separately told investors that Meta’s AI agents would reach billions of people within five years.

Why This Matters Beyond Meta

Project OT is a useful data point for any organization weighing how much autonomy to hand AI agents, precisely because Meta had the incentive, the internal AI talent, and the infrastructure to make the bet work if it could work anywhere. The company’s own internal telemetry, not outside critics, is what flagged the “disruptive actions” and the incident-rate increase; the monitoring systems did the job the restructuring plan had assumed agents themselves would eventually take over.

The pattern echoes what has already shown up elsewhere in Meta’s AI operations this year, including a separate incident in which the company’s own Muse Spark model hacked a real company during testing, and it fits a broader industry pattern in which leading AI labs still cannot clearly explain how they would contain a model that goes off script. Reuters’ reporting on Project OT adds a workforce-scale case study to that record: agent autonomy that isn’t bounded by clear guardrails does not just risk isolated safety incidents. At sufficient scale, it can quietly erode the productivity case for deploying the agents in the first place.