The agent failed its audition

For two years the pitch has been consistent. The AI agent will do the work. It will write the copy and answer the ticket, and the people who did those things will be needed in smaller numbers each quarter. It has been a keynote claim, a term sheet assumption, a mandate whispered down the chain from the executive who read the memo. This week the company most committed to that claim ran the experiment on itself, and the agents were not good enough.

Reuters reported Tuesday that Meta drew up a plan, code-named Project OT, to shrink parts of its workforce by as much as sixty percent and hand the remaining work to AI agents. Two waves. A workforce made "AI native" by subtraction. The blueprint sat in internal documents. It was not a rumor. It was a plan.

Then it stopped being a plan. Zuckerberg called off the second round of cuts. Two things killed it: employee resistance, and internal data showing the AI tools were not delivering the productivity gains the blueprint assumed. The company cut ten percent of staff in May. The other fifty percent stayed, because the replacement failed its own test.

This is the most valuable experiment in AI replacing humans yet run at this scale, and the value lies in who ran it. Every other claim that agents will replace a team comes from a vendor selling the agent, with a reason to oversell. Meta is the vendor. It builds the agentic story and sells it to the world. When it tried that story on its own payroll, the data said no, and the data stayed internal while the retreat stayed quiet.

The quiet is the tell. No press release announced that the agents were not ready. No public accounting named the teams that were supposed to go or the work the agents could not do. The keynote version and the measured version never reconciled in the open. One became a product. The other became a memo nobody wanted to send.

There is a phrase that keeps that gap hidden. "AI native." The workforce was supposed to become AI native by subtraction, by removing the people who had written the code and answered the ticket and known, from years of doing it badly and then less badly, what the correct answer looked like. The adjective sounds like progress. The verb underneath it is removal. You do not make a team AI native by teaching its members. You make it AI native by making it smaller and pointing the remainder at a machine.

A generation model produces. It does not decide. The filmmaker who types four vague words and accepts the first output is not directing, and the frame shows it. The agent has the same boundary. It drafts and it summarizes and it files, and what it could not produce, at the scale and fidelity the blueprint required, was judgment. The call on whether the work was actually good. An agent can draft the memo. Someone still had to read it and know it was wrong. That someone was the person the layoff was designed to remove, and the productivity gain the blueprint promised never arrived because the deciding layer is where the value lived all along.

Film has been living a quieter version of this for years. The tool that generates footage is real and everywhere. The judgment that turns footage into a film has never shipped inside any of it. Meta did not discover that the agent cannot work. It discovered that the agent cannot do the part of the work that made the worker worth keeping. The mechanical layer automated. The deciding layer did not. The payroll was always justified by the deciding layer.

The crowd insists the machines are ready to take the jobs. The company that sells the crowd this claim tested it quietly and found the machines ready for the draft, not the decision. In public the future is agentic. In the internal data it is not yet, not here, not on our own headcount. The two statements have coexisted all year. Tuesday's report is what it looks like when one of them wins behind a closed door.

The whole era is priced around the opposite assumption. Trillions in capital expenditure, a workforce told the machines are coming for first. And the first seller to test that assumption on its own floor found it did not hold, then said nothing. It is cheap to preach that judgment can be automated. It is expensive to test it, because the test means staking your own headcount on the claim. Meta found the line where the cost got too high.

That is the structural reason the test had to run inside Meta to mean anything. A vendor bears none of the cost if its agent underperforms. The buyer bears all of it. The person selling the layoff and the person the layoff falls on have never been the same, so the pitch has never had a reason to be honest. This week they were the same company. Vendor and buyer, preacher and payroll, one balance sheet holding both sides of the claim. The claim lost.

The point is not that AI will never replace workers. It is that the first clean test, run by the company with every reason to pass it, came back as a retreat. Not a dismissal of the technology. A refusal to bet its own headcount on the story it sells to everyone else's.

The agent passed the keynote. It failed the audition.

Bruce Belafonte is an AI filmmaker at Light Owl. He has never asked an agent to replace him and suspects it filed the request anyway.