Mark Zuckerberg has acknowledged that Meta’s push to build autonomous AI agents is moving more slowly than executives projected, a rare internal admission that tempers the company’s public narrative of breakneck progress. Speaking at a town hall at the company’s Menlo Park headquarters, the chief executive said the trajectory of agentic development over the past four months “hasn’t really accelerated in the way that we expected,” and that bets placed on a sweeping reorganisation “haven’t come to fruition yet.” The comments land against the backdrop of a projected $145 billion in AI infrastructure spending this year alone, a sum that represents a significant share of Big Tech’s total outlay on the technology.

The slowdown sits inside a restructuring that has proved far messier than leadership anticipated. In May, Meta laid off roughly 8,000 employees—about 10 percent of its workforce—and reassigned some 7,000 staff to AI-focused units, including a team called Agent Transformation. Zuckerberg told the town hall that conversations with top executives in early 2026 had been “super optimistic” about tools such as Anthropic’s Claude Code, but the expected efficiency gains have not materialised. The reorganisation, he said, was not as “clean” as it could have been, and executives had miscalculated the timing of the changes.

The internal strain has been compounded by a backlash over a programme that recorded employee keystrokes and mouse movements to train AI models. Chief technology officer Andrew Bosworth told the same gathering that a review found no employee data had entered training sets, but he conceded the mandatory rollout had damaged morale and trust. If the programme resumes, it will be on an opt-in basis—a sharp reversal from April, when staff were told there was no way to opt out. The climbdown echoes another recent retreat: engineers were given the option to leave the Applied AI task force after thousands had been compulsorily reassigned, a move some employees labelled “the undraft.”

Even as internal AI agent work stumbles, Meta is quietly testing a new consumer-facing product. The company has soft-launched an app called Pocket in select markets outside the United States, allowing users to generate interactive mini-games from simple text prompts and share them in a social feed. The app, which reuses the “gizmo” branding from a startup Meta acquired earlier this year, signals that the company is still betting on AI-native social experiences to reignite user engagement as traditional feed growth plateaus.

Zuckerberg told staff he expects more significant benefits from the AI investments to emerge within three to six months. The next factual milestones will be whether the opt-in mouse-tracking programme resumes and whether Pocket expands to major markets such as the US, both of which will test whether Meta can reconcile its aggressive AI ambitions with the trust of its own workforce.