Reuters: Zuckerberg's AI Push Stumbles, Code Crisis Halts Layoffs

Original author: Keluode (Claude), TechFlow

TechFlow (深潮) Introduction: A Reuters investigation published on Aug. 26 revealed that Meta internally modeled an "AI-native" reorganization in early 2026: some teams could be cut by up to 60%, with agents taking over daily work and smaller "high-density talent" groups overseeing virtual workers. On the night of May 19, Zuckerberg halted the second wave originally planned for November. The next day, the company still laid off about 10% of staff. Internal data showed code changes rose 220% year-over-year, but features delivered to users rose only 36%; major technical and safety incidents rose 40% year-over-year, and employee favorability dropped from 74% to 55%.

Reuters published an investigative report on Aug. 26: Zuckerberg had set an organizational overhaul early in the year to replace a large amount of routine human labor with AI, codenamed Project OT (Organization Transformation). In the most aggressive scenario, many teams could be reduced to 40% of their original size. The plan was to be implemented in two waves, in May and November. On the night of May 19, he halted planning for the second wave.

Reuters reviewed dozens of internal documents, posts, and recordings, and spoke with more than 20 people familiar with the matter. Meta acknowledged the year-long project but described it as "scenario planning," emphasizing that it was not intended to cut 60% of the company's workforce and that "not every scenario was pursued."

Hawaii Closed-Door Meeting Sets Tone, Some Teams Modeled for 60% Cuts

In January 2026, Zuckerberg held the annual leadership offsite with his core team at his Hawaii estate. According to sources cited by Reuters, the direction set there was to make Meta an "AI-native" company. An internal planning document stated that AI would take over much of the daily work of thousands of employees, with virtual workers supervised by smaller, more "talent-dense" human teams within the company.

In the scenario planning, multiple teams could be cut by up to 60%. Some people would be transferred to new units, others let go. A human resources executive estimated in another internal document that the scale of this round of cuts could match or exceed the company's roughly 25% layoff in 2023. The reorganization was planned in two waves: the first cut in May, another in November, along with closing open positions and removing those deemed low performers.

These scale details have not been previously reported publicly. Meta's written response to Reuters made clear: the most aggressive scenarios did include reducing some teams by up to 60%, with paths including both layoffs and transfers; the company never intended to cut 60% of the entire group; several key business units were not in scope. The statement read: "As part of our company reorganization earlier this year, we asked some teams to do scenario planning to assess the potential impact of transfers, closing open roles, and layoffs." "This ultimately resulted in thousands of employees moving to several newly formed teams to work on priority areas, which has been publicly reported. We did not pursue every scenario modeled—and never assumed all would be implemented."

Reuters could not confirm a single trigger for Zuckerberg's decision that night, and the company declined to make him available for an interview.

Teams Shrink from Dozens to Three to Five, "Builders" Replace Old Roles

After ChatGPT launched in late 2022, Silicon Valley management became fascinated with "agents": they should not just answer questions, but place orders, book flights, and write applications themselves. Meta internally codified this philosophy as "AI-native"—one Project OT document stated: "AI-ready tools and agents collaborate, workflows are automated, and new projects default to AI-first." Externally, the company also prepared to sell business agents capable of booking and closing deals to other enterprises, launching an enterprise version in June.

Head of Product Naomi Gleit spent "a considerable amount of time" in the Singapore office last year. She told Reuters in June that the approach there "inspired some teams in California and New York," and many changes were "bottom-up." Chief Data Officer Alex Schultz also visited Asia last year, and according to three sources, the group admired how local startups were redesigning their org structures around AI.

An earlier pilot took place in July 2025. Ime Archibong, a long-time VP of product management, announced in an internal post that his team would first create five "small engineering squads," each with two to three engineers and one designer, equipped with AI tools, using four-week sprints to replace product cycles of about six months. He used a basketball analogy: "In basketball, embracing the fast break lets you take more shots and better shots. Our expectation with AI tools is the same: they let us explore more ideas at lower cost and higher fidelity." The next sentence: "This isn't just more fun—in an AI-first era, we believe this is the winning strategy."

In October 2025, people on his team posted the "AI-Native Playbook," a guide for other groups on "how to cross over": traditional roles of product designers and engineers disappear, squad members are uniformly renamed "builders"; middle management is removed; squads report directly to a senior unit lead; daily priorities are handled by "agent-assisted analysis."

Early this year, Zuckerberg pushed Project OT into implementation, requiring management to change structures. By June, at least 11 units (including engineering and research) had switched to the squad model. An internal notice said performance ratings and promotions would use a "village approach": unit leads managing 30 to 50 people make the decisions, assisted by HR and unnamed "AI systems"; squad leads only handle day-to-day matters without formal management authority. One employee assigned to lead a squad wrote on an internal message board: "I didn't go through management training, and I don't have access to ratings and management tools." Meta emphasized to Reuters that ratings and promotions "have been and continue to be human decisions, not AI."

Traditional product groups typically have 10 to 20 people: seven to fourteen engineers, plus one each for product, design, data science, user experience research, and data engineering. The AI-native squads in the design documents have only 3 to 5 people. HR was simultaneously looking for "irreplaceable talent" and "10x producers"—theoretically one person plus AI equals ten people, and the savings from layoffs would be used to give them raises.

Code Up 220%, Features Delivered to Users Up Only 36%

The core of the plan was autonomous agents. Internal data did not support that core.

Chief Technology Officer Andrew Bosworth wrote in an internal post in early June: code changes on internal software platforms and infrastructure used daily by employees increased 220% year-over-year; changes that actually became new features or upgrades delivered to Meta users increased only 36% year-over-year. Infrastructure teams had warned as early as March that the flood of AI-generated code was causing "reliability alerts." Another internal post in April was more blunt: unconstrained agents were executing "large-scale, destructive operations that humans would be unlikely to attempt."

The results were written in the same internal posts. Major technical and safety incidents (service outages, potential data leaks, etc.) rose 40% year-over-year, and time spent by employees on incident response increased 70% year-over-year. In early June, attackers also used Meta's AI customer service bot to gain access to a number of high-profile Instagram accounts, which Reuters identified as including the now-defunct Obama White House account. Meta declined to comment on these internal failure data.

At an internal town hall in early July, Zuckerberg admitted the timeline was miscalculated. In a recording heard by Reuters, he said "at least over the past four months, the trajectory of agent development has not accelerated as we expected," and the bet on the company's new structure "has not yet paid off." He also said he hoped to see more tangible returns in the next three to six months. At the same event, he admitted the reorganization was "not smooth enough," and that when planning in January and February, senior leaders' concern was "we are not adapting fast enough."

Investors were asking a different question. Meta's capital expenditure guidance for this year had been raised to $130 billion to $145 billion, mainly for AI chips and infrastructure. The Reuters investigation wrote "at least $130 billion," and cited LSEG estimates that this spending would consume the company's 2026 operating cash flow. Second-quarter earnings subsequently confirmed the pressure: free cash flow fell 91% year-over-year.

Employees See Monitoring and Transfers as Training AI to Replace Them, Favorability Drops from 74% to 55%

When the May 20 cut landed, employees no longer believed "transformation" was just about changing tools. Reuters first reported on March 13 that the company was discussing layoffs of 20% or more, and in April wrote about the first wave of about 10% on May 20 and possible additional cuts within the year. Sources told Reuters that the reporting disrupted management's original communication rhythm.

Alongside layoffs, a large number of engineers were pulled into new units like Applied AI Engineering to work on training data. Earlier Reuters reporting stated that the May global layoffs were about 10%, roughly 8,000 people, with about 7,000 more transferred to new teams related to AI workflows. At the end of the second quarter, the company had about 75,500 employees. Some of those forcibly transferred called themselves "conscripts" internally. On U.S. devices, keyboard and mouse tracking was also required to collect keystrokes and mouse trajectories to train agents that can mimic human computer operation. Employees posted angry messages and sarcastic jokes on the internal Workplace platform; some replied to executives' posts with elephant images, implying layoffs were the elephant in the room that couldn't be discussed. Earlier reports said more than 1,600 people signed a petition against the monitoring.

In the semi-annual Pulse survey, employee favorability dropped from 74% to 55%. Labor organizing was also heating up. On the night of May 19, planning for the second wave was halted. The next day, the first wave proceeded as scheduled. Zuckerberg then told remaining employees that he "does not expect any more company-wide layoffs this year," hoping to provide more "stability." The company later paused keyboard and mouse tracking, allowed some of those pulled away to return to their original teams, and increased budgets for travel, team building, and snacks. Employees on internal channels focused on his repeated qualifiers—"company-wide" and "this year"—speculating that headcount would continue to drop through team eliminations or performance-based removals, or that the next cut would be pushed to next year.

In June, Zuckerberg shifted the external narrative back to "betting on people" in an internal post: "We are the only large company focused on empowering people and putting this new technology in the hands of billions of users, rather than primarily using it to automate jobs."

The Reuters investigation included this statement after the plan had already been scaled back. Currently, Meta's official position on this investigation remains that "not all scenarios were implemented."