BOTTOM LINE
Meta just quietly scaled back its most aggressive plan to replace workers with AI agents. Not because the tech failed outright. Because folding agents into a real organization turned out to be its own unsolved problem, one a $135 billion budget couldn’t buy past.
The Plan Was Simple. The Execution Wasn’t.
Internally it was called Project OT, according to a Reuters investigation. The pitch: shrink some teams by up to 60%. Replace most of the work with AI agents. Keep a small pod of skilled humans managing them.
A first wave of cuts went through. A second, bigger wave was set for November.
Zuckerberg pulled it hours before it was supposed to launch.
The Math That Broke It
Here’s the part worth sitting with. Code changes pushed through Meta’s AI platforms jumped 220% year over year. Actual user-facing features shipped? Up just 36%.
Technical incidents rose 40%. Fixing them ate 70% more employee time than before.
That’s the real lesson buried in the numbers. Agents burn tokens writing the code. Then they burn more tokens, and more human hours, cleaning up what they got wrong. The cleanup is costing more than the code is worth.
By July, Zuckerberg admitted agentic development “hasn’t really accelerated in the way that we expected.” That’s about as blunt as a CEO gets in public.
The Part No Spreadsheet Predicted
Meta had rolled out software tracking keystrokes and mouse clicks, sold internally as productivity monitoring. Employees eventually figured out the data was training their own replacements.
Sentiment scores cratered. Internal boards filled with anger. People started organizing.
You can model incident rates and token spend. You can’t model a workforce that realizes it’s watching itself get automated in real time.
Not a Failure. A Recalibration.
Meta didn’t kill the project. It throttled the aggressiveness back while it figures out the actual ratio of agents to humans that works.
And that’s the detail I keep coming back to: Meta wasn’t the only company betting big that agents were ready to run the show, and if the company with the deepest pockets and its own frontier lab can’t shortcut this, that tells you the blend isn’t a solved problem anywhere yet. It’s not a Meta problem. It’s an everyone problem.
So Is the Doomsday Scenario Wrong?
Worth being clear about what this doesn’t mean. It doesn’t mean AI’s effect on jobs isn’t real, or isn’t worth watching closely. The infrastructure spending hasn’t slowed. The ambition hasn’t changed.
But the version of the future getting sold to us, mass layoffs by Christmas, entire departments swapped for agents overnight, doesn’t survive contact with an actual org chart. Even inside the company most incentivized to make it work fast, the transition looks less like a light switch and more like a long, uneven negotiation between what the technology can do and what an organization can absorb without breaking itself.
That’s not comforting exactly. But it’s a more honest timeline than the one currently being sold as inevitable.
