Meta has put its AI-centered reorganization plan on hold and withdrawn the second round of layoffs it had scheduled for November. The plan's goal was to turn the company into a so-called "AI-native company," in which AI agents handle a large share of employees' daily work and fewer people oversee the results. During the trial period, code output rose 220%, but outages rose 40% over the same period. Further layoffs, which had been under review on a scale of several thousand people, were halted as well.
An AI agent is software that, once a person gives it a single instruction, breaks the work into steps on its own and pushes it through to execution. A chatbot only produces answers, while an agent takes on writing, fixing and shipping code. Meta's idea was to raise this tool from a personal aid to a unit of organizational structure. The idea was discussed at a meeting held in January at the Hawaii home of Meta CEO Mark Zuckerberg.
In the results, only one indicator moved much. A 220% rise in code output means the total volume of what the organization produces grew more than threefold. But outages rose 40%, and the measures of actual feature conversion and of stability fell short of expectations. The gap between how much was built and how much works has widened.
This gap is also an old problem in software development. Writing code and making that code hold up in a production environment are different jobs, and the latter requires context and a clear line of responsibility. Agents handle the first part quickly, but the added burden of review and recovery on the second part comes back to people. Cutting headcount while increasing output narrows the point where that burden lands.
Meta reviewed a plan to cut the headcount of some teams by up to 60%. Which units that scale would apply to was not settled at the review stage. The fact that the reorganization stopped at that point shows that setting a replacement ratio in advance and fitting the organization to it could not withstand the actual measurements.
The broader industry trend points the same way. An analysis by Andreessen Horowitz presented a finding that AI agents are consuming more AI tokens than people are. That means agents are already handling a considerable workload, but there is no guarantee that consumption amounts to results. Meta's case comes close to the first measured evidence that the two have to be looked at separately.
Silicon Valley companies are redrawing their organizations into a form in which agents and people work together, rather than stripping out staff. AI use has also been found to be growing in the legal field. The focus is shifting from whom to replace to whom to place where.
Whether Meta has postponed the second round of layoffs or abandoned it entirely has not been disclosed. But seen through a single developer's day, the change started some time ago. The screen opened in the morning is stacked with code the developer did not write, and it takes time to sort out which of it to trust and which to roll back. What organizations will have to count from now on is less how many people remain than how many people can handle that sorting work.
