At Meta, workers train their replacements
Thirty days before cutting 8,000 employees, Meta began recording the keystrokes of those still working
THE FACEBOOK era was built on data users produced for free in exchange for a feed. The agent era, Meta revealed on Tuesday, will be built on data its own employees produce for a salary.
On April 21st, the Model Capability Initiative began quietly installing itself on US employees' laptops. MCI, per an internal memo first surfaced by Reuters, captures mouse movements, clicks, and keystrokes across work-related applications, and takes occasional snapshots of whatever is on screen. The stated purpose, as described by a staff AI researcher writing to the Meta SuperIntelligence Labs channel, is to teach Meta's models how humans "actually use" computers — specifically dropdown menus and keyboard shortcuts, the mundane micro-decisions agents have been conspicuously bad at replicating. Andrew Bosworth, Meta's chief technology officer, framed the broader effort as the training bed for a newly rebranded Agent Transformation Accelerator, itself the third renaming of the initiative in four months. By Tuesday afternoon, internal backlash had been sharp enough to prompt a follow-up memo acknowledging, per messages seen by Reuters, "a lot of concern about this."
Smoke and keystrokes
The surveillance framing will dominate the coverage, and Meta has made that easy: a spokesperson's assurance that data will not be used for performance review is paired with a conspicuous absence of detail about what counts as "sensitive" and therefore excluded. The more interesting story is downstream. Meta's defense — that agents cannot be built without behavioral data — is accurate. What has gone under-discussed is how expensive that data is to acquire, and how neatly Meta's workforce solves the problem.
Academic groups have spent the past eighteen months trying, and largely failing, to assemble behavioral-trajectory datasets at the scale agent training demands. OpenCUA, an open-source framework out of Hong Kong, painstakingly collected 22,600 recorded task demonstrations across Windows, macOS, and Ubuntu — a project that consumed months and dozens of annotators. A Shanghai Jiao Tong paper published last year squeezed a 141% performance jump out of just 312 high-quality human trajectories, confirming that trajectories, not model architecture, are the scarce input. Scale AI and Surge AI, the incumbent labelers, price their work as enterprise-bespoke; Surge's reported 2024 revenue alone topped $1 billion, most of it from major AI labs commissioning exactly the kind of human-generated demonstration data agent training now requires. The public internet, well-mapped for text models, contains almost nothing of what an agent needs to learn. Watching someone navigate Salesforce at 11 a.m. on a Tuesday is apparently not a YouTube genre.
Meta has roughly 79,000 employees, the majority of whom spend eight hours a day clicking through the exact kinds of enterprise software agents must eventually replicate. The marginal cost of each recorded trajectory, once MCI is deployed, is effectively zero. A billion-dollar labeling contract pales against a payroll already on payroll.
The timing is uncomfortable. Meta will begin its first wave of layoffs on May 20th, cutting roughly 8,000 people — about 10% of the global workforce — with further reductions planned for the second half of the year. Company spokespeople insist MCI data will not drive performance reviews, and there is no evidence that employees being recorded are disproportionately among those being cut. The pattern is structural, not individual. Each engineer inside Meta's newly formed Applied AI Engineering division — a unit explicitly chartered to build agents handling "large portions of product development" — is generating training data for the very agents Meta is building to handle adjacent roles. Oracle cut nearly 18% of its workforce this year to fund $156 billion of AI infrastructure; Amazon cut 16,000 in January; 44% of US hiring managers now cite AI as a primary driver of planned cuts. What Meta has done differently is put the workforce-as-data-source arrangement in writing, in a memo signed by its chief technology officer.
The Bosworth memo ends on an aspiration: that agents will eventually "automatically see where we felt the need to intervene so they can be better next time." Next time, of course, is the point. For the 8,000 Meta employees about to be notified of their departure, the open question is how much of their work has already been recorded, and how useful it has been. Meta's next training run will know. They will not.
