Mark Zuckerberg’s message to employees this week cut against the tone of Meta Platforms’ public AI push. In an internal meeting, the chief executive said the company’s AI agent development had not accelerated the way he expected over the past four months, according to people who attended.
He added that the restructuring of the company’s AI efforts, which included large-scale layoffs, could have been handled more cleanly, and that Meta’s bet on a new technical architecture had not yet paid off, the people said. The remarks, first reported by Chinese financial media, offered a rare view of internal doubt at a company whose executives have spoken confidently about agentic AI.
Meta has poured money into artificial intelligence over the past three years. It has built large compute clusters, trained a family of open-source models under the Llama name and shipped AI assistants across its apps. The company has told investors that AI would reshape advertising, feed recommendation and, eventually, the way people delegate tasks to software.
Zuckerberg’s comments suggest the last part is proving harder than the marketing. Turning a conversational model into a reliable agent, one that can be trusted to book a trip, manage an inbox or negotiate on a user’s behalf, has been slower than the industry’s rhetoric implied. Executives and engineers across the sector have made the same point privately for months.
The admission is notable because Meta has been among the most aggressive corporate backers of agents. Its research arm has published work on agent frameworks and planning models, and the company has described AI assistants embedded in its messaging apps as an early form of the technology. Zuckerberg has said in public appearances that agentic AI would be a defining theme of the coming years.
The reorganization he referenced has been disruptive inside the company. Meta has moved teams between its AI research groups and its product organizations, and it has cut roles in areas it judged less central to its AI plans. Layoffs have hit parts of the company that were expanded during the earlier platform boom, and some departing employees have described the process as abrupt.
Zuckerberg said the restructuring could have been cleaner, language that suggested dissatisfaction with execution rather than with strategy. He also said the company’s bet on a new architecture had not yet shown results, the people said. Meta has said little publicly about the specifics of its model infrastructure, and people familiar with the company’s work said the reference pointed to changes in how its models are trained and served.
Analysts said the remarks matter for investors because they temper expectations. Meta’s stock has traded at levels that assume AI will generate new products and revenue within a couple of years. If the company’s own chief executive is telling employees that progress is slower than planned, the timeline for those products may stretch.
The wider industry faces the same gap. OpenAI, Google and a field of startups have all shipped agent features, and all have hit the same wall: models that chat fluently still struggle with long, multi-step tasks where mistakes compound. Reliability, not capability, has become the binding constraint, engineers in the field said.
Meta’s AI spending has become a point of tension with Wall Street. The company has raised its capital expenditure guidance repeatedly, and investors have mostly accepted the increases while revenue growth remained strong. Zuckerberg’s comments could sharpen the question of how long that patience lasts, because they suggest the payoff from agents is further out than executives have implied in earnings calls.
Inside the company, the meeting added to a stretch of turbulence. Meta has reorganized its AI units several times, and some engineers have said on internal forums that priorities shift faster than projects can ship. The chief executive’s acknowledgment that the new architecture had not yet shown results validated a complaint that had been circulating among staff, people familiar with the discussions said.
For Meta, the stakes are double. Its advertising business funds the AI spending, and investors want to see the models pay for themselves. At the same time, rivals are converting their AI research into consumer products, and a slower agent roadmap leaves Meta exposed in the one category it has staked out for the future.
People who attended the meeting said Zuckerberg did not announce new deadlines or a change of plan. He framed the pace as a fact to be managed rather than a failure to be punished, and he repeated his view that AI would eventually justify the investment. But the tone, several people said, was more sober than the company’s public statements.
Meta declined to comment on the meeting. The gap between a chatbot and a tool that finishes real work is wider than outsiders assume, and the company’s own chief executive now appears to agree.
