Zuckerberg admits Meta's AI agents are moving slower than expected: the hype meets reality

By Reuters · July 2, 2026.

According to a recording of an internal town hall heard by Reuters, Mark Zuckerberg told employees that the development of AI agents over the past four months "has not accelerated as we expected." He added that the company's reorganization—which included major job cuts—"was not as clean" as it could have been, and that the bets made on the new structure "have not yet paid off." Meta projects it will spend up to $145 billion on AI infrastructure this year, a significant portion of the more than $700 billion that Big Tech will collectively devote to the technology. Zuckerberg himself said he expects more tangible benefits from that investment over the next three to six months. At the same meeting, CTO Andrew Bosworth separately addressed a security incident involving the mouse-movement tracking software the company uses to train AI, confirming that no employee data was leaked and that the program, if reactivated, will be opt-in.

The news matters less for what it says about Meta specifically than for what it reveals about the real state of the AI agent race across the industry. Over the past two years the dominant narrative has been that autonomous agents—capable of planning, executing complex tasks and operating with minimal human oversight—were just around the corner, ready to reshape entire business processes. That the CEO of one of the companies pouring the most money into that bet is the one publicly admitting, before his own employees, that progress fell short of expectations is a valuable data point: it is rare to get this kind of unfiltered honesty from inside a frontier lab.

Broadly, the AI sector is caught between two speeds: that of benchmarks and demos, which advance with spectacular rapidity, and that of real integration into productive workflows, which is far slower and bumpier. Agents that work well in a controlled environment stumble against the variability of the real world: ambiguous contexts, legacy systems, the need for near-perfect reliability when autonomy is delegated to them. Meta is not the exception; it is probably the norm that other giants prefer not to admit out loud. The difference is that here we have the verbatim quote.

There is also a human dimension that should not be overlooked: the layoffs tied to that reorganization have already been carried out, with the social cost that implies for those affected, while the promised benefits—the reason that in theory justified the cuts—have yet to materialize according to Zuckerberg himself. It is exactly the pattern we have been flagging in our coverage of AI and employment: companies act first on the promise of efficiency, and the evidence that such efficiency arrived tends to lag months or years behind. Those who lose their jobs do not recover that time even if the technology ultimately delivers.

Our reading is that admissions of this kind, far from being bad news for the underlying case for AI, are a sign of the sector's maturity: clearly distinguishing between what already works and what is still aspiration is exactly the exercise we advocate in the face of overblown marketing. That Meta continues to commit $145 billion to infrastructure despite results below expectations suggests the industry is betting on a horizon of several years, not quarters, before autonomous agents generate consistent returns. That is compatible with our underlying thesis: the transition toward an AI-driven economy of abundance will be neither linear nor fast in the short term—there will be painful restructurings, promises unmet within the projected timeline and capital burned on bets that take longer than announced—but the direction of the investment, and the persistence in it even after acknowledging friction, is consistent with a genuine leap in capability that is still to come, not with a bubble deflating.

The mouse-movement tracking episode, though secondary in the report, adds an uncomfortable layer: it is a reminder that part of the data infrastructure with which these systems are trained has been built with questionable internal surveillance methods, and that the response—switching to opt-in after the controversy—comes only after public pressure. It is a reminder that the race for AI agents is not fought solely in the model labs, but also in the everyday decisions about which data is collected, from whom and with what consent.