Zuckerberg: Meta AI Agent Push Hasn't Accelerated as Hoped

Mark Zuckerberg admitting, on the record to his own staff, that the last four months of AI agent work has not paid off is the kind of quiet signal worth pausing on. Meta's CEO reportedly told an internal town hall on Thursday that AI agent development over the last four months hasn't 'accelerated in the way we expected,' according to Reuters, and that leadership's bets on the recent AI reorganization 'haven't come to fruition.' He also said the reorg wasn't as 'clean' as they would have liked it to be.

The context is what makes it interesting. Meta has raised its 2026 capex guidance to $125-$145 billion, and Zuckerberg told employees plainly that the upcoming layoffs are driven by that capital reallocation, not by AI making internal work more efficient. As The Next Web summarized the town hall, the cuts are about paying for GPUs and data centres. The company is spending like agents will land soon, and staffing like the payoff is still uncertain.

Why this matters if you don't run Meta: agent capability is the load-bearing assumption under a lot of 2026 AI economics. If the operator with the biggest checkbook and the freshest reorganization, under Chief AI Officer Alexandr Wang's Superintelligence Labs, is saying in private that the curve is flatter than planned, that is a real data point about how hard the last mile of agent reliability actually is.

The honest caveats: this is a leaked internal talk, sourced from a Reuters-heard recording, and 'hasn't accelerated the way we expected' is not the same as 'isn't working.' What the reporting doesn't spell out is which agent products slipped, what the internal benchmark was, or how Wang's group is being measured against it.

The forward read is that rivals with less exposure to a single mega-reorg, and enterprise buyers evaluating agent claims from any vendor, just got a useful piece of ground truth from the party least incentivized to share it.