At an internal town hall, the CEO said agent development hasn't sped up the way leadership predicted, and conceded the reorganization behind the bet was messier than it should have been.
Mark Zuckerberg has told Meta employees that the company's push into AI agents has moved slower than leadership banked on when it rewrote its organizational chart at the start of the year, a candid admission that lands months after Meta cut roughly a tenth of its workforce to fund the effort.
Speaking at an internal town hall on Thursday, according to a recording described by Reuters, the Meta chief executive said the pace of agent development over the previous four months had not picked up the way executives forecast. He did not stop at a status report. Zuckerberg conceded the reorganization behind the bet was not as "clean" as it should have been, and that his team had misjudged the timing of the changes.
Agents sit at the center of the industry's current pitch. They are systems meant to carry out multi-step work on a user's behalf rather than answer a single prompt, and Meta has made them a pillar of its long-term strategy. Zuckerberg's verdict on the reshuffle built to chase them was plain. The bets on the new structure, he said, "haven't come to fruition yet."
A Sharp Turn From January's Confidence
The tone is a reversal. Eight months ago Meta was moving with the urgency of a company convinced it was behind.
Zuckerberg said the trajectory of agent development over the last four months had not accelerated as leadership anticipated, a line that reads far apart from the optimism inside Menlo Park at the start of 2026. He offered employees a timeline for patience, telling them Meta should begin to see more meaningful returns from its AI spending within the next three to six months. It is the kind of horizon that has a way of sliding forward, and some in the room have heard versions of it before.
A Meta spokesperson declined to comment on the remarks.
One in Five Employees Caught in the Overhaul
The restructuring Zuckerberg was defending was not small. In May, Meta laid off about 10% of its global workforce, roughly 8,000 people by Bloomberg's count, and moved another 7,000 or so into AI-focused groups, one of them named Agent Transformation. Taken together, the cuts and reassignments touched close to a fifth of the company.
The changes bruised morale. Employees pushed back on the scale and the speed, and the reassignments arrived as a signal that entire teams were being pointed at a target the technology had not yet reached. Sharpening the resentment, some of the reshuffling coincided with Meta paying enormous packages to poach outside AI talent, a gap in treatment that current staff noticed.
Zuckerberg tried to close that wound in May, telling employees he did not expect further companywide layoffs this year. Not everyone believed him.
His concession this week, that the reorganization was messier than it needed to be and mistimed, is as close to an apology as the workforce is likely to get. He and other executives have spent recent weeks walking back pieces of the earlier plan without abandoning its direction, a balancing act meant to steady the company while admitting the original blueprint was flawed.
Claude Code and the Fear of Falling Behind
Zuckerberg was direct about where the original conviction came from. When leadership began mapping the restructuring in January and February, he said, the worry among senior people was speed. Conversations "with our top people," in his words, circled a fear that Meta was not moving fast enough to keep up.
Part of that anxiety had a name. Executives were "super optimistic" about coding tools like Claude Code, the software agent from rival lab Anthropic that had built a following among developers for handling real programming work with limited supervision. Tools in that mold convinced Meta's leadership that a broad wave of capable agents was arriving fast, and that the company needed to reshape itself to ride it.
The wave has been slower and choppier than that read suggested. The tools do help, particularly with code. What has not shown up on schedule is the sweeping, dependable autonomy that Meta reorganized thousands of jobs to capture.
A $145 Billion Bill Coming Due
None of this is happening cheaply. Meta is on track to spend as much as $145 billion on AI infrastructure this year, a figure that sits inside a Big Tech buildout topping $700 billion across the largest US technology companies. The company has separately committed to $600 billion in US infrastructure spending through 2028.
That spending has changed how Wall Street reacts to Meta. Investors who once cheered aggressive AI budgets have started to flinch. When Meta added billions to its capital plans alongside a third of a year in revenue growth, the stock fell rather than rose, part of a broader shift in which the market now scrutinizes the size of the checks more than it applauds the ambition behind them.
The pressure Zuckerberg faces is arithmetic. Data centers and chips do not wait for the payoff, and the power bills land on the same schedule, with depreciation on that hardware hitting the books whether or not the agents arrive. A three-to-six-month promise of returns is not just a message to employees. It is a message to shareholders watching the distance between what Meta is spending and what its AI has so far produced.
Inside a Lab That Keeps Reorganizing
The agent stumble is the latest chapter in a stretch of near-constant reshuffling inside Meta's AI operation. Last year Zuckerberg spent $14.3 billion for a stake in Scale AI and installed its founder, Alexandr Wang, then 28, atop a new umbrella group called Meta Superintelligence Labs. The move was meant to end the company's catch-up posture against OpenAI and Google.
What followed was churn. Meta dissolved its earlier AGI Foundations group and split the lab into four teams. By March it had stood up a parallel Applied AI Engineering organization under Reality Labs veteran Maher Saba, reporting to CTO Andrew Bosworth. The two-track setup exposed a rift over direction, with Wang's side pushing frontier research toward models meant to rival the best from competitors, while Bosworth and product chief Chris Cox favored wringing consumer products out of Meta's own data.
The lab did ship. Its first model, Muse Spark, arrived in April and now powers the Meta AI assistant, with an upgrade promising stronger coding and agent skills on the way.
Meta also put a commercial face on the agent push. In June it launched Meta Business Agent at its Conversations conference in London, a customer-facing assistant that can book appointments and close sales inside WhatsApp and Messenger, with Instagram added at rollout. More than a million businesses were already using earlier chatbot versions, and the launch nudged Meta's stock up about 4% on the day. The product works. The internal transformation meant to stand behind it is what Zuckerberg now says has lagged.
The Whole Industry Is Learning Agents Are Hard
Meta is not alone in discovering that agents are harder to ship than to announce. By most estimates only a small slice of organizations, somewhere around 11% to 14%, have moved agentic systems into real production, and the ones succeeding tend to be the ones that moved carefully rather than quickly.
The sticking point is rarely raw intelligence. Teams building with these systems most often cite integration with existing software as their central obstacle, followed by the reliability problems that pile up when an agent chains many steps across many tools. One weak link breaks the chain, and the failures compound where nobody can see them.
There is a serious bull case underneath the doubt. Goldman-linked forecasts have token consumption rising more than twentyfold by 2030, driven largely by enterprise agents, which is precisely why hyperscalers are willing to spend ahead of visible revenue. If agents do embed themselves in daily business workflows, today's infrastructure turns into recurring demand rather than a stranded bet.
Skepticism has still climbed to some of the industry's most prominent figures. Former Intel chief Pat Gelsinger has said flatly that AI spending carries bubble-like features, and OpenAI's Sam Altman has acknowledged a bubble even while expanding the company most responsible for inflating it. The capital is real. The revenue is real. The distance between the two is what has investors uneasy.
Against that backdrop, Zuckerberg's town-hall confession reads less like a Meta-specific misstep and more like one large company saying out loud what many are learning in private.
Bosworth and the Surveillance Program That Backfired
The same meeting surfaced a second bruise. Bosworth used it to address the fallout from a controversial employee-monitoring program that Meta paused in June after a data-security failure.
The tool, known internally as the Model Capability Initiative, was pushed onto US employees' laptops in April. It logged staff keystrokes and mouse clicks and grabbed occasional screenshots. That behavior was fed into Meta's models as training data, the idea being to teach AI systems how human workers actually move through software.
Workers hated it. More than 1,600 signed a petition against the program, and Meta's later addition of a 30-minute pause button mostly underlined how constant the watching otherwise was.
Then it leaked. A permissions misconfiguration, not an outside attack, left the collected data readable across roughly 45,000 internal tables. According to documents reviewed by Reuters and other outlets, the exposed material included AI prompts, employees' private conversations, performance information and, in some reporting, records as sensitive as personal tax and medical data. Meta paused the program on June 22 and opened a review.
The legal exposure is real, particularly in Europe. Keystroke logging and screen capture of identifiable employees runs straight into the General Data Protection Regulation, which sets a high bar for processing personal data and treats workplace consent as shaky given the power imbalance between an employer and its staff. The irony was not lost on employees either, who were being asked to train the systems Meta is building to automate work like theirs, only to have that harvested data mishandled.
At the town hall, Bosworth said the review had found no employee data was included in AI training. He also floated a concession. If the program returns after the review, he told staff, it would run on an opt-in basis. People who are comfortable can contribute to what he called a "great human survey," and those who are not, he said, would not have to.
That is a notable climbdown. When Meta first installed the software in April, Bosworth had told employees there was no way to opt out.
Comments
Join the discussion and share your perspective.