Meta Tried to Replace Workers With AI Agents. Why 'Project OT' Fell Apart

In late August 2026, multiple outlets reported that Meta had abandoned an internal restructuring plan, codenamed Project OT for "Organization Transformation," that would have reshaped the company around AI agents doing work previously done by employees. The plan reportedly envisioned cutting team sizes by as much as 60 percent and running the remaining work through small "pods" of humans overseeing fleets of agents. What makes this worth analyzing is not that a plan was scrapped, companies shelve plans constantly, but the specific reason it reportedly failed, which is one of the cleaner real-world tests of a claim the whole industry is making.

The plan, as reported

Project OT is described as an attempt to make Meta "AI native": rather than treating AI as a tool employees use, the company would restructure so that AI agents carried much of the daily work, with engineers, designers and product managers converging into general-purpose "builder" roles supervising them. The reporting says the cuts were planned in two waves, one earlier in 2026 and a second later in the year, and that the second wave was scrapped after Meta had already reduced its workforce by roughly 10 percent in May.

The goal was not AI as a tool employees use. It was AI as the thing that does the work, with people supervising the machines.
On the reported ambition of Project OT

Set aside whether the plan was wise. As a controlled experiment, it is unusually informative, because Meta is not a company short on AI talent, models or compute. If autonomous agents were going to replace large fractions of skilled knowledge work anywhere in 2026, a company with Meta's resources trying hard to make it happen is close to a best case. The reported outcome is therefore a meaningful signal about where the technology actually is, not just about one company's management.

The number that explains the reversal

According to the reporting, the plan ran into a wall that was measured, not merely felt. Internal data reportedly showed that code changes rose about 220 percent year over year as engineers leaned on AI tools, while the changes that actually reached users as new or upgraded features rose just 36 percent. In other words, the machines produced vastly more output, and only a small fraction of that output turned into product people could use.

More code, not proportionally more product

Reported year-over-year change at Meta. Output (green) surged; the share that reached users as shipped features (grey) rose far less. The gap is the review-and-judgment bottleneck AI did not remove.

Source: figures attributed to internal Meta data in reporting on Project OT, late August 2026. Reported, not independently audited.

That gap is the analytically important part, and it echoes a pattern documented well beyond Meta. Generating code, text or designs is the step AI has made dramatically cheaper. Turning that raw output into something correct, coherent, safe and worth shipping is the step it has not. When you remove the humans who did the second step, you do not get proportionally more product; you get a pile of generated work that still needs judgment applied to it, and a bottleneck that simply moved rather than disappeared. A 220 percent rise in code against a 36 percent rise in shipped features is a precise picture of throughput outrunning outcomes.

A person in a suit signing a colorful wall covered in handwritten messages at Meta's headquarters — Reporting says an employee backlash, including a sharp drop in sentiment after tracking software was installed to train AI, compounded the plan's technical problems. Photo: U.S. Department of the Treasury, public domain.

The human failure ran alongside the technical one

The reversal was not purely a productivity story. Reporting says Meta installed tracking software on US employees' computers to capture how they worked, as training data for the agents meant to replace them, and that this contributed to a roughly 19-point drop in employee-sentiment scores. A workforce asked to train its own replacements, while being monitored to do so, is unlikely to produce the enthusiastic, high-quality work that such a transition would need.

Asking people to train the agents meant to replace them, while tracking how they work, is a hard way to get the cooperation the transition requires.
On the reported employee backlash

That interaction between the technical and the human is easy to miss and important to name. Even if the agents had been more capable, a demoralized workforce degrades exactly the tacit knowledge and judgment that the "builder pod" model depended on. The two failures reinforced each other: agents that could not yet carry the work, and people with little reason to help them get there. Neither alone might have killed the plan; together they made it unworkable.

What this is, and is not, evidence for

It would be a mistake to read Project OT's reversal as proof that AI cannot change knowledge work. That is not what the reporting shows. Meta still reportedly kept the roughly 10 percent of cuts it made earlier, and the 220 percent rise in code changes is itself evidence that AI is now deeply embedded in how the work gets done. The technology clearly changed the throughput of the work. What it did not do, on this timeline, is let a small group of supervisors reliably convert that throughput into finished product without the specialists in between.

The honest reading is a specific one. As of 2026, autonomous agents can amplify skilled workers dramatically, and cannot yet replace the judgment layer that turns their output into something shippable, at least not at the scale and reliability a company like Meta needs to bet its structure on. That is a narrower and more useful conclusion than either "AI replaces everyone" or "AI changes nothing," and it is the one this episode actually supports.

What it means for teams building with AI

For anyone deploying AI inside a real organization rather than reading about it, the lesson is to measure the right thing. Output metrics, lines of code, drafts generated, tickets closed by an agent, are easy to celebrate and, on the evidence here, easy to inflate without moving the outcome that matters. The bottleneck moved to review, integration and judgment, and the teams that get value are the ones that resource that layer rather than assume the agent removed it.

Part of getting there is matching the tool to the task instead of forcing one model to do everything and then cleaning up the mess. Different models are better at different steps of that pipeline, and the frontier keeps shifting between them. Infrastructure that stays model-agnostic, the approach platforms like Metir take by routing work across models from multiple labs, is one way to keep output quality high enough that more of it actually ships, which is precisely the gap Project OT ran into. The point is not more generated work; it is more work that clears the judgment bar.

The bottom line

Project OT's reported collapse is a rare, well-resourced test of the "agents replace knowledge workers" thesis, and the result is instructive precisely because it is mixed. AI made Meta's engineers far more productive by one measure and barely moved the measure that counts, shipped product, while an attempt to remove the humans in between backfired both technically and culturally. Read without spin, it is neither a refutation of AI's impact nor a vindication of the replace-everyone plan. It is evidence that in 2026 the technology amplifies skilled work faster than it can substitute for the judgment that finishes it, which is a more durable finding than any headline about layoffs.

Sources:

Image credits

Hero: Meta CEO Mark Zuckerberg speaking at a public event. Wikimedia Commons, by Anthony Quintano, licensed CC BY 2.0.
Meta headquarters: a visitor signing the wall at Facebook's Menlo Park headquarters. Wikimedia Commons, U.S. Department of the Treasury, public domain.