Meta's AI Workforce Plan Imploded as Agents Failed and Employees Revolted

A Reuters investigation published Wednesday reveals that Meta secretly planned to replace up to 60% of positions in certain teams with AI agents — a program code-named Project OT, for Organization Transformation — but abandoned the second wave of the plan on the evening of May 19, hours before it was set to execute, after the agents failed to deliver measurable productivity gains. The collapse, reconstructed from scores of internal documents and recordings, offers the most detailed account yet of what happens when a major technology company tries to replace employees with AI workers and discovers the technology isn't ready.

The investigation appears three weeks after Zuckerberg published a 6,500-word essay predicting AI would create "an abundance of jobs." The Reuters documents show that his own experiment to demonstrate that transition had already stalled.

Project OT: An AI-Native Company Built at a January Retreat

Zuckerberg and his senior leadership team developed the initiative during their annual January 2026 retreat at his Hawaii compound. The plan was built around a structural conviction: that conventional software teams of 10 to 20 specialists were inefficient by design, and that a smaller number of highly capable people — the plan called them "talent-dense" groups — could supervise AI agents handling the underlying work.

In practical terms, Project OT envisioned rebuilding Meta around small "pods" of three to five generalist employees, replacing specialized titles for product designers and engineers with a single flexible designation: "builder." Pod leads would report to senior unit heads who might oversee 30 to 50 people, collapsing the management layers in between. Daily priorities would be set not by managers but by what an internal document called "agent-assisted analysis." Four-week sprints would replace the six-month product planning cycles that had governed Meta's engineering process for years.

Senior executives including Chief Data Officer Alex Schultz and Head of Product Naomi Gleit had spent time observing how Asian technology startups organized themselves around AI tools before returning to develop the approach at Meta. The initiative was designated high-priority, given a codename, and given a two-wave implementation schedule — May 2026 and November 2026.

Read more: New research gives AI agents a principled formula for deciding when to seek context

Meta confirmed the scenarios were explored but stressed the 60% figure applied only to specific teams, not to the company's overall headcount. Project OT combined team-level cuts with eliminated open positions, redeployments, and removal of low performers — a four-part approach designed to shrink headcount without a single company-wide number.

The Agents Couldn't Do the Work

By the time the May 20 layoffs began — 8,000 employees notified, plus 6,000 open positions canceled and 7,000 workers transferred into newly formed AI units — the internal evidence that the plan's central premise was failing had already accumulated.

The most striking data point came from a June post by CTO Andrew Bosworth, reviewed by Reuters: code changes to Meta's software platforms and infrastructure had increased 220% year over year as employees adopted AI coding tools. Changes that resulted in new or upgraded features actually reaching users rose only 36%. The gap is a direct measurement of the production problem at the core of enterprise agentic AI: AI tools generated vastly more code, but testing, review, integration with legacy systems, and deployment — the last mile — remained almost entirely human work, and that work was under pressure from the volume surge.

The reliability picture was worse. Internal posts reviewed by Reuters described "reliability warning signs" tied to the AI coding surge as early as March. An April post warned that unchecked AI agents were performing "large-scale, disruptive actions that humans are unlikely to execute." Major technical and security incidents — including service outages and possible data leaks — rose 40% from the prior year, while the time employees spent responding to those incidents jumped 70%. Rather than freeing engineers to do more productive work, the AI expansion was generating a new and growing category of cleanup work.

Meta's internal half-year Pulse survey captured what employees made of all this. Favorable sentiment dropped from 74% to 55%. Employees began posting angry messages, jokes, and images of elephants — a reference to the layoffs as "the elephant in the room" — on Meta's internal network. Some of them turned to memes printed on flyers and posted in Meta's US offices mocking the company's AI ambitions.

The Surveillance Program That Became a Flashpoint

Running parallel to Project OT was the initiative that became its most visible liability. In April 2026, Meta quietly deployed software to US employees' work laptops under a program it called the Model Capability Initiative, or MCI. The software logged keystrokes, recorded mouse movements, tracked click locations, and captured periodic screenshots — all on company-issued devices, with no opt-out option. Meta framed MCI internally as a way to collect "real examples of how people actually use computers" to train AI agents capable of performing office tasks autonomously.

Employees understood it differently. They connected MCI to Project OT and concluded they were being surveilled in order to generate the training data that would be used to build their own replacements. More than 1,600 employees signed an internal petition calling for the program to be canceled outright. UK-based staff began a formal union organizing drive. A recording of Zuckerberg reportedly defending the program as a competitive necessity — and acknowledging it would not be in the company's "strategic interest" to fully explain it to employees — circulated internally.

A security review subsequently found that the data captured by MCI, including AI prompts, private conversations, and employee performance records, had been accessible across thousands of internal data tables far beyond the intended audience. Meta paused the program. CTO Andrew Bosworth later told employees the review concluded no employee data had been used in AI model training. If MCI resumes, he said, participation would be voluntary — a significant reversal from the no-opt-out deployment.

The regulatory exposure is geographically asymmetric. European employees were exempt from the start: the EU's General Data Protection Regulation prohibits monitoring employee computer activity without explicit consent, and the EU AI Act's high-risk system provisions, which came into force on August 2, 2026, include specific obligations on AI systems that interact with employees. US employees on company-issued devices have no equivalent federal protection; California, Connecticut, and Delaware require written notice, but not opt-outs. Yale law professor Ifeoma Ajunwa has noted that US law imposes no federal limit on worker surveillance. A discrimination lawsuit filed in July in the US District Court for the Northern District of California alleges that behavioral telemetry scoring used in the May layoffs couldn't account for workers on protected medical, parental, or disability leave — potentially the first legal test of AI-assisted layoff selection under the FMLA.

Why AI Agents Fail at the Production Layer

Meta's experience is consistent with what independent research has documented about enterprise agentic AI at scale. Understanding why requires a brief account of how these systems actually work.

An AI agent operates in a continuous perception-plan-act-evaluate loop. It receives a high-level goal, decomposes it into sub-tasks, selects and calls external tools — APIs, databases, code execution environments — evaluates whether the result advances the goal, and then iterates. Unlike a chatbot that processes a single input and produces a single output, an agent is supposed to manage entire workflows autonomously: receive a request, search for data, draft a document, schedule a follow-up, log the result, without a human approving each step.

The architecture works reliably in controlled conditions. Production environments introduce failure modes that demos do not surface. Context windows degrade under sustained load when long task chains push against token limits. Tool-call schemas become inconsistent when multiple users hit the same system simultaneously, causing agents to misinterpret API responses. Most critically, errors compound across multi-step task chains: a misclassification in step two of a ten-step workflow does not produce a minor error at step ten; it produces a broken workflow, often in a way that requires human review to diagnose. Meta's internal April 2026 post described agents executing "large-scale, disruptive actions that humans are unlikely to execute" — precisely the error-compounding behavior that production-grade agentic systems have not yet reliably solved.

Industry data confirms the gap is structural rather than Meta-specific. Roughly 31% of enterprises had at least one AI agent in production as of Q1 2026, per Gartner, McKinsey, and S&P Global research, with banking leading at 47% and healthcare at 18%. A 2026 McKinsey follow-up found fewer than 10% of enterprises that experimented with agents had scaled one to deliver measurable value, and a widely cited MIT study found 95% of enterprise AI pilots showed no measurable financial return. Gartner projects more than 40% of agentic AI projects will be abandoned by the end of 2027, primarily due to unclear ROI, escalating costs, and inadequate governance.

The Forward-Deployed Alternative Meta Bypassed

What the industry has concluded, and what the AWS and Microsoft announcements the week of Zuckerberg's July 2 town hall made explicit, is that production AI agent deployment cannot be achieved through organizational restructuring alone. It requires intensive human engineering support embedded inside the organizations where agents are supposed to work.

Amazon Web Services, on June 30, announced a new forward-deployed engineering organization backed by $1 billion, deploying thousands of engineers directly inside client companies to build and operationalize agentic systems — compressing timelines, AWS said, "from months to days." The following day, July 2, Microsoft announced the Microsoft Frontier Company: a $2.5 billion operating unit staffed by approximately 6,000 engineers, built explicitly to embed technical staff at enterprise client sites and make AI agents work in production environments. Anthropic and OpenAI have launched structurally similar programs.

The forward-deployed model was pioneered by Palantir, which discovered that data infrastructure only delivers value when experts are present inside client organizations to adapt it to specific workflows. The AI industry has arrived at the same conclusion through the same mechanism: failed production deployments that exposed the gap between general-purpose model capability and the context-specific engineering required to make agents reliable.

Meta's Project OT represented a different theory: that internal reorganization alone could compress the journey from model capability to production output. Zuckerberg acknowledged the outcome at his July 2 internal town hall, telling employees in a recording obtained by Reuters that "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 bets on the new structure "haven't come to fruition yet."

The Gap Between the Essay and the Data

On August 10, 2026, Zuckerberg published "The Future Is for Everyone," a 6,500-word essay arguing that broadly distributed superintelligent AI would produce an "abundance of jobs" rather than eliminate them, and positioning Meta's open-source model strategy as a safeguard against AI power concentrating in too few hands.

The essay appeared 39 days after the July 2 town hall admission and 17 days before the Reuters investigation, which drew on internal documents spanning precisely the period of failure the essay addressed. The juxtaposition — public optimism about AI-driven job abundance written against a private record of AI agents disrupting rather than replacing human work — is now part of the public record.

Meta did proceed with the May layoffs: 8,000 employees, $1.18 billion in severance recorded in Q2, headcount at 75,472 as of June 30. Zuckerberg's June 12 memo acknowledged "mistakes" and pledged no further company-wide layoffs for the rest of 2026. Thousands of employees were redeployed rather than let go, and Project OT's vision of smaller, AI-supervised teams persists in modified form.

The real test of the Project OT thesis does not arrive with the Reuters investigation. It arrives in Q4 2026, when Zuckerberg's three-to-six-month window from the July 2 town hall elapses and investors can ask directly whether the restructuring has produced the productivity returns that justified it. With up to $145 billion in 2026 capital expenditure and an industry-wide pattern of production AI agent deployments stalling at the same structural bottleneck, the question is not whether Meta will try again — it is whether the agents will have advanced far enough by then to close the gap that Project OT opened.