Zuckerberg Concedes Meta's AI Agent Push Is Behind Schedule, Six Months After the Reorg Meant to Fix It

Mark Zuckerberg rebuilt his company around a bet on AI agents. On July 2, standing in front of his own employees, he admitted the bet has not paid off on schedule.

"The trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected," Zuckerberg said at an internal Meta town hall, according to an audio recording obtained by Reuters and reported by outlets including PYMNTS and The Decoder. The company's wagers on its new structure, he added, "haven't come to fruition yet."

The venue matters. Not an earnings call, not a keynote — a closed-door all-hands that leaked — which is why the language is blunter than anything Zuckerberg has said publicly about Meta Superintelligence Labs since he created it. A Meta spokesperson declined to comment on the recording.

What was actually said, and what it cost

The admission arrives on top of one of the most expensive corporate restructurings in tech history. Meta plans to spend as much as $145 billion on AI infrastructure in 2026 — a single-company figure that represents a meaningful slice of the more than $700 billion Big Tech is collectively pouring into compute this year.

The human cost is equally concrete. In May, Meta laid off roughly 10 percent of its global workforce — about 8,000 people out of a headcount near 78,000 — while simultaneously moving approximately 7,000 employees onto newly formed AI teams. The stated logic was a swap: trade headcount for compute, fund the data centers, and recoup the lost labor through AI-assisted workflows.

Zuckerberg told the town hall that the planning for this began in January and February, when senior leaders worried Meta was not moving fast enough. At the time, he said, executives were "super optimistic" about coding tools like Anthropic's Claude Code. He conceded the reorganization was not as "clean" as it could have been and that leadership had misjudged the timing.

One caveat on the timeline in this story's own headline: the six-month gap runs from when the restructuring was designed, not when it landed. The plan was locked in January and February; the layoffs did not execute until May. The four months Zuckerberg described as disappointing — roughly March through June — overlap with a period when much of the reorg was still being implemented.

The dissent came from inside the building

The most interesting wrinkle is that Meta's own AI chief did not agree with the framing.

At the same town hall, Alexandr Wang — the Scale AI founder Zuckerberg installed atop Meta Superintelligence Labs — struck a markedly more bullish tone. Meta's next model, code-named "Watermelon," has caught up with OpenAI's GPT-5.5, Wang said, citing benchmarks he did not specify. "Watermelon, our next model after Avocado, is currently in training," Wang said, per Business Insider. "Watermelon uses an order of magnitude more compute than Avocado" — the internal name for Muse Spark, the frontier model Meta shipped in April with respectable benchmark scores that nonetheless trailed OpenAI and Anthropic.

Wang then went further on X, arguing that Zuckerberg had been describing the pace of the entire industry rather than Meta specifically. That reading is contested: the quote as reported by Reuters is embedded in a discussion of Meta's own restructuring bets, and Zuckerberg's surrounding remarks about a reorg that was not "clean" are self-evidently about Meta. Readers should treat Wang's reinterpretation as spin until a fuller transcript surfaces.

Analysis: the gap between agent hype and agent shipping

Zuckerberg's candor lands awkwardly against the industry's public forecasts. Gartner projects that 40 percent of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5 percent in 2025 — an eightfold jump in roughly eighteen months.

"AI agents will evolve rapidly, progressing from task and application specific agents to agentic ecosystems," Anushree Verma, Senior Director Analyst at Gartner, said in the firm's forecast. "This shift will transform enterprise applications from tools supporting individual productivity into platforms enabling seamless autonomous collaboration and dynamic workflow orchestration."

Both things can be true, and that is the real story. Embedding a task-specific agent into an existing enterprise app — a summarizer in a CRM, a triage bot in a ticketing queue — is a shipping problem, and vendors are clearly solving it. Building general-purpose agents reliable enough to justify a $145 billion infrastructure bill and 8,000 eliminated jobs is a research problem, and research does not respond to reorg charts.

Meta's specific error looks like a category confusion between the two. The company restructured in January and February on the assumption that agentic coding tools would compound productivity fast enough to make a smaller workforce more productive than a larger one. Four months later, the compounding had not arrived on schedule — but the layoffs had.

There is a hedge in Meta's favor. Zuckerberg framed the shortfall as a timing miss rather than a directional one, and told staff he expects more significant benefits from Meta's AI investments within the next three to six months. Meta is also reportedly building a cloud business to sell excess AI compute to outside customers, per Bloomberg, a revenue path independent of its own agent timeline.

What to watch

Three markers. First, whether Watermelon ships and whether independent benchmarks support Wang's GPT-5.5 parity claim, or whether it becomes the second consecutive model to post good numbers and lose the comparison. Second, Meta's next earnings call: if Zuckerberg repeats any version of the town hall language in front of analysts, the capex guidance becomes the story. Third, Gartner's 40 percent figure — if enterprise agent embedding hits that mark while Meta's own agents slip, the lesson will not be that agents were overhyped. It will be that the boring integration work was the winnable part all along.