On August 26, Reuters published an investigation into Project OT, short for Organization Transformation. It was Meta’s plan to have AI agents take over daily work from thousands of employees, with smaller teams supervising them. Reuters based the report on internal documents, posts, and recordings, and on interviews with more than 20 people.
Meta confirmed the plan existed and that it was structured in two waves. It also confirmed that its scenarios included cutting some teams by up to 60 percent. The company says that figure applied only to certain teams, that several major units were excluded, and that it never intended to cut 60 percent of the company.
According to Reuters, the first wave went ahead in May. About 8,000 jobs were cut, roughly 10 percent of staff, and 7,000 people moved into AI roles. Mark Zuckerberg called off the second wave. Meta told Reuters it never assumed it would act on every scenario.
Reuters also reported internal figures that Meta declined to comment on. Code changes to internal platforms rose 220 percent year on year. Changes that reached users as new or improved features rose 36 percent. Major technical and security incidents rose 40 percent, and time spent fighting them rose 70 percent. In July, Zuckerberg told staff that agentic development “hasn’t really accelerated in the way that we expected.”
Output went up. Output that reached users went up far less. Incidents went up too.
Activity is not the same as results
Measured by code produced, the agents looked like a success. Measured by what reached users, the gain was a fraction of that. Measured by incidents, there was a new cost that a headcount plan has to absorb. The same rollout tells three different stories depending on which number gets reported upward.
The figures are internal and reported, not confirmed by Meta. Most organizations won’t have Meta’s engineering depth or its budget, which makes the lesson cheaper to learn from someone else.
Before you plan headcount around agents
- Measure what reaches customers or residents, not what the tools produce.
- Track incidents and the time spent fixing them alongside any productivity number.
- Pilot with one team, with a written plan for reversing course.
- Keep the people who would fix what agents break until you know what breaks.
- Keep the cost case separate from the capability case. Budget decisions shouldn’t run ahead of the pilot data.
The plan assumed the agents would work. The incident log said otherwise.
Sources