Meta's 'OT Project' aimed to replace thousands of employees with AI agents, but it backfired: code volume surged, safety incidents rose by 40%, and remaining employees spent 70% more time fixing problems. Zuckerberg halted the second round of layoffs, but the company is now facing a mess three times worse than before.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Meta's 'OT Project' envisioned replacing thousands of employees with AI agents, but the result was a 40% increase in incidents, a 70% increase in firefighting time, and an emergency halt to the second round of layoffs by Zuckerberg. This failure highlights the gap between AI capability predictions and actual production.
Early this year, Zuckerberg convened core executives at his private estate in Hawaii, proposing a radical vision to replace thousands of employees with AI agents, leaving only a handful to supervise. This evolved into the internal 'OT Project' (Organizational Transformation), with its 'AI-native' definition including tool-agent interaction, workflow automation, and selling AI agents to third parties.
The OT Project stemmed from a pilot program by product executive Ime Archibong last year: five small tech teams (each with two or three engineers and a designer) used AI tools to develop prototypes in four-week sprints. Meta also published the 'AI-Native Playbook,' detailing how to cut middle management and use 'agent-assisted analysis' to prioritize tasks.
Executives envisioned one employee doing the work of a former team. In the most aggressive scenarios, some teams would be cut by 60%, and HR executives said the workforce could shrink by about 25% or more. The plan originally called for two rounds of restructuring: the first in May and the second in November, with savings partly used to reward high performers.
On May 20, the first round of layoffs took place, cutting about 10% of employees and moving a similar number to new projects. However, on the evening of May 19, just hours before the layoffs, Zuckerberg canceled the November round. Reuters could not determine the specific reason for his change of direction.
Meta responded that it had canceled the OT Project before determining specific layoff numbers, emphasizing that scenario planning was not fully implemented. But employees found that AI didn't do their work; it created more.
According to an internal post by CTO Andrew Bosworth in early June, code changes to internal software platforms and infrastructure increased 220% year-over-year, but only 36% of changes translated into new features or upgrades reaching users. Much of the AI-written code became 'noise.'
Internal documents show AI agents performed 'large-scale destructive operations that humans could hardly have executed,' leading to a 40% increase in major technical and safety incidents compared to the previous year. Employees spent up to 70% more time resolving issues.
On May 30, Instagram suffered its most embarrassing security outage: attackers, using only usernames, filed customer service appeals with fake locations, tricking Meta's customer service AI into sending verification codes to attacker-controlled emails, enabling zero-authentication password resets. Obama's official account was also compromised. Insiders say the direct cause was AI-generated code plus AI-reviewed code, and the trust and safety team lost about half its staff due to reassignments. The next day, Meta's CISO Guy Rosen announced his resignation.
Employee anger erupted, with internal systems flooded by negative comments, and morale scores dropped from 74% to 55%. One employee cursed during an internal livestream, saying they had become 'the company's servant.' In May 2026, the number of Meta engineers registering for new job searches increased significantly year-over-year, with some core engineers deciding to leave due to 'lack of autonomy.'
Zuckerberg admitted at an internal meeting in July: 'At least in the past four months, the trajectory of agent technology hasn't really accelerated as we expected.' This marked a reflection by leadership on AI capability predictions.
Sanchit Vir Gogia, chief analyst at Greyhound Research, noted: 'Meta believed in an AI capability prediction that hasn't been realized in production. This is a different failure from over-trusting AI.' He argued Meta counted predicted capabilities as actual capacity, recommending verifying AI reliability before expanding permissions and removing human controls.
This failure sparked industry reflection on the path of 'replacing humans with AI.' One netizen commented: 'They glimpsed a future possibility and thought it had arrived, so they started laying off.' AI capabilities are unevenly distributed and cannot quickly replace human employees at this stage.
Meta ultimately faces an organization with damaged morale, technical debt, and a clearer understanding of AI's boundaries. Engineers are dealing with a mess three times more chaotic than before: code volume tripled, safety incidents up 40%, and firefighting time up 70%.
This article is based on reports from Reuters, Ars Technica, and InfoQ. Details such as internal documents and posts are as reported by media and have not been independently confirmed by Meta. Meta officially responded that it canceled the OT Project but did not confirm specific layoff plans. Some data (e.g., code volume increase, incident increase) comes from internal documents or executive posts and is at the source_claim level.
The failure of Meta's OT Project shows that counting AI capability predictions as production capacity is dangerous. AI excels at certain tasks, but its overall capabilities are uneven and cannot quickly replace human employees. Companies should gradually verify AI reliability rather than aggressively laying off staff.
Primary report
Primary source