Does more code proportionally create more product? No

How the biggest experiment in replacing people with AI played out

In theory, there is no difference between theory and practice. In practice, there is. The first results of replacing people with AI tools at real scale have arrived, and not from just anywhere, but from Meta. In theory, everything added up: the Project OT plan, agents taking over the work of thousands of employees, some teams shrinking by 60%, two waves of layoffs. In practice, it went differently: an employee revolt, an Instagram breach through an AI support bot, the second wave cancelled hours before the first one began, and, as a closing note, a promise to improve the snacks in office kitchens. Reuters reconstructed the whole story from dozens of internal documents, posts and recordings.

In January, Zuckerberg gathered his top executives for the annual offsite at his Hawaii compound. That is where Project OT, Organization Transformation, was launched. The internal documents reviewed by Reuters describe the vision as follows: agents take over the daily work of thousands of employees, overseen by small «talent dense» groups of people. Scenario planning included cutting some teams by as much as 60%. An HR executive estimated the scale as comparable to or larger than the purge of 2022-2023, when the company cut roughly 25% of its workforce. The plan was split into two waves, May and November. A separate line item covered selling agents externally: Meta intended to offer other companies agents for scheduling meetings and closing sales.

The vision has its own history. After ChatGPT arrived, Silicon Valley developed a philosophy of «AI native» companies: rather than embedding AI into existing processes, you build the organization around it from scratch, so that agents and tools interact with each other and workflows are automated by default. Meta’s leaders went to see what this looks like in real life: in Asia they were impressed by startups with org structures built around AI, and the head of product said openly that the Singapore office had inspired teams in California. The company commissioned its own study of how AI startups are structured and launched pilots. In July 2025, VP of product Ime Archibong assembled five pods of two or three engineers and a designer equipped with AI tools, running four-week sprints instead of six-month planning cycles. He explained the idea through basketball: a fast break gives you more shots and better positions, and AI tools would likewise let teams test more ideas cheaper and better, which in the AI era is the winning strategy. In October, an internal document called the «AI Native Playbook» scaled the approach: engineer and designer titles disappear, everyone becomes a «builder», middle management is removed, and agent-assisted analysis helps set daily priorities.

By June, at least 11 units had moved to pods. A «village» model was introduced: one Org Lead per 30-50 people, making ratings and promotion decisions with support from HR and unspecified «AI systems», while a Pod Lead runs the team day to day but holds no formal authority. One person appointed to that role wrote on the internal network that he had been given neither manager training nor access to evaluation tools. In parallel, HR built an «Irreplaceable Talent» search tool with a «10X Performer» archetype, and the savings from layoffs were earmarked for packages for AI engineering stars. Some engineers were transferred to a unit called Applied AI Engineering, where they wrote coding problems to train Meta’s models. In internal posts, people called the work mind-numbing and menial. In some engineering units, headcount had fallen 30% by the end of May.

Then the company started losing people before the layoffs even began. In March, Reuters reported on the coming cuts before even vice presidents had been briefed. Leadership went silent, and managers were handed talking points saying that roles would «evolve» because of AI. In April, it emerged that tracking software recording keystrokes and mouse clicks was being installed on employee devices in the US, to teach agents how to work at a computer. Employees put two and two together: they were training their own replacements. The internal network filled with angry posts, elephant pictures appeared under leadership announcements, the elephant in the room, flyers with a petition against the tracking hung in office bathrooms, and people publicly clashed with CTO Bosworth. When Zuckerberg announced an AI initiative for small businesses, an employee acidly compared him to Prometheus bringing fire to mankind. The loyalty index in the semi-annual survey dropped from 74% to 55%, and union organizing efforts gained momentum.

And against this backdrop came the key numbers. In early June, Bosworth wrote in an internal post that the volume of code changes on internal platforms had grown 220% year over year. Changes that reached users as new or updated features had grown by 36%.

The mechanics of the gap are simple. Code generation is one part of building a product, and before AI it was not the narrowest one. Written code has to be reviewed, tested, shipped, and then kept in production for years: monitored, fixed, updated, with someone answering for the consequences. Agents accelerated only the first step. The throughput of the rest of the pipeline stayed the same, the same reviewers, the same on-call engineers, the same infrastructure. When three times as much code is poured into a system with unchanged throughput, the difference does not disappear. It accumulates in review queues, in unverified changes, in technical debt, and it surfaces as incidents.

Meta’s internal data shows exactly this picture, step by step. As early as March, infrastructure teams warned of reliability warning signs, and in April an internal post recorded that unchecked agents were taking large-scale disruptive actions that a human would not have taken. Major incidents, including service outages and potential data leaks, grew by 40%, and the time engineers spent firefighting them by 70%. In June, the problem went public: hackers used an AI support bot to gain access to major Instagram accounts, including the Obama-era White House account.

On the evening of May 19, hours before the first wave was due to start, Zuckerberg gathered his inner circle and called off the November wave. What exactly made him change his mind, Reuters could not establish. On the morning of May 20, 10% of the workforce was laid off, around 8,000 people, after which he wrote to employees that he expected no other company-wide layoffs this year and wanted to give people stability. Then came the repair work: mouse tracking was paused, some engineers were allowed to return to their old teams, the CFO took charge of perks, down to promises of better snacks in office kitchens. In July, at an internal meeting, Zuckerberg admitted that agentic technologies were accelerating more slowly than he had expected and promised results within three to six months.

Externally, meanwhile, a «betting on people» campaign rolled out, with videos about Meta betting on its people. Zuckerberg published an essay, «The Future is for Everyone», predicting an abundance of jobs in the future, it is just that companies will get smaller while there will be more of them. At the same time, in internal communications he holds carefully to the words «company-wide» and «this year», and employees read this as a door left open for targeted cuts. The pressure has not gone anywhere: AI capex this year is around 130 billion dollars, and analysts expect it to consume all of the operating cash flow of 2026.

The belief that «more code means more product» is one of the most persistent misconceptions about vibe coding. Meta tested it on itself, with the best models and a budget nobody else has. Code grew by 220%, product by 36%. And even those 36% cannot be counted as pure gain. From them you have to subtract the 40% rise in incidents and the 70% rise in engineers’ firefighting time, which means service downtime, money, and team attention pulled away from development. Add the collapse of employee trust, the cancelled programs and the months of repairing relationships, and the final sign of the effect becomes hard to call. Acceleration definitely happened. Acceleration of what is an open question.

The main takeaway: Meta ran the experiment for the entire industry. At its own expense, one the rest of the industry could not have afforded in either money or nerve, and now all of us have numbers instead of assumptions. The one who does nothing makes no mistakes. Experiments must be run, and large companies run them just as startups do. Some work out, some do not, and that is the essence of entrepreneurship: nobody knows the right answers from the start, even with all the money and resources in the world. You have to try and you have to build.

AI will, of course, change our lives and our work. But not everything, not immediately, and not everywhere.