Mark Zuckerberg told Meta employees on Thursday that the company’s AI agents have not progressed as quickly as he expected, four months after a restructuring that was supposed to speed things up.
“The kind of trajectory of the agentic development over at least the last four months hasn’t really accelerated in the way that we expected,” he said at an internal town hall, according to Reuters.
The admission lands awkwardly against the scale of what Meta has already spent chasing that acceleration. The company is projected to spend up to $145 billion on AI infrastructure this year, part of a restructuring that included cutting roughly 8,000 jobs in May while simultaneously moving thousands of staff onto AI-focused teams.
Zuckerberg said that when the reorganisation was being planned in January and February, executives were “super optimistic” about coding tools such as Anthropic’s Claude Code, and had expected that optimism to translate into faster agentic progress across Meta’s own products.
It hasn’t, by his own account. He told staff the bets made during that period “haven’t come to fruition yet,” and acknowledged the reorganisation itself wasn’t as clean as it could have been, with the timing of the changes partly driven by fear that Meta “weren’t going to move fast enough to adapt.”
That fear was not abstract. Reuters reported that around 7,000 employees were reassigned into AI roles the same week roughly 10% of the global workforce was cut, a reshuffle Zuckerberg had previously framed to staff as a matter of capital spending priorities rather than AI itself replacing jobs.
That framing traces back to an earlier town hall in May, where he told employees the company runs on two cost centres, compute and people, and that headcount would keep bending toward the former.
Despite the shortfall, Zuckerberg struck a forward-looking note on the timeline, telling employees he expects Meta to see “more significant benefits” from its AI investments within the next three to six months.
He did not specify which products or teams would deliver those benefits, and a Meta spokesperson declined to comment when contacted by Reuters.
The same town hall covered a separate, thornier issue. Meta’s chief technology officer, Andrew Bosworth, told staff that an internal review had found no employee data from a paused mouse tracking and keystroke monitoring tool had been used to train Meta’s AI models.
That tool, part of what the company calls its Model Capability Initiative, was rolled out in April without an opt-out option, prompting internal pushback before it was paused. Bosworth said the programme may now resume, but only on an opt-in basis, a reversal from its original design.
Meta’s struggles with agentic AI sit alongside broader strain inside its AI organisation. Engineers inside the company’s Applied AI unit have described the unit’s working conditions as punishing since the restructuring pulled staff in from other parts of the business.
Zuckerberg’s comments suggest the culture strain and the technical delays are, at minimum, running on parallel tracks, whatever the actual causal relationship between them turns out to be.
Meta is not alone in treating agents as the next competitive front. Rivals including OpenAI, Google and Anthropic have all pushed agentic products this year, betting that AI systems capable of completing multi-step tasks without constant human prompting will be the feature that justifies the industry’s spending.
Zuckerberg’s comments are among the more candid admissions from a major AI lab that the underlying technology has not moved as fast as the spending implies.
Meta reports second-quarter earnings later this month, which will give investors their first chance to ask Zuckerberg directly how the gap between capital outlay and agentic progress squares with the company’s spending guidance for the rest of the year.
Get the TNW newsletter
Get the most important tech news in your inbox each week.
