Somewhere in Menlo Park, an engineer is writing a React component in VS Code right now. Every keystroke, every dropdown click, every file navigation — it's all being recorded. Not by malware. By their employer.

Meta announced Monday that it's rolling out an internal program called the Model Capability Initiative. The tool installs on U.S. employees' work computers and captures mouse movements, clicks, keystrokes, and periodic screenshots. The stated goal: teach AI models how humans actually use computers so the company can build agents that do it instead.

What the Tracker Sees

The monitored applications include Gmail, GChat, VS Code, and an internal tool called Metamate. Those represent the full workflow surface of a knowledge worker: email, messaging, coding, and internal collaboration. The software records granular interaction patterns — dropdown navigation, keyboard shortcut preferences, file-switching habits, application transitions — plus periodic screen captures for context.

The company says this data feeds exclusively into AI training pipelines and won't touch performance reviews. Whether you believe that depends on how much you trust a company that's been fined billions for privacy violations.

"Primarily Do the Work"

CTO Andrew Bosworth didn't mince words about the endgame. The vision involves building agents that "primarily do the work" while humans "direct, review and help them improve."

Not "assist with the work." Primarily do the work. When your CTO says that and then installs a keylogger to make it happen, the subtext isn't really sub anymore.

Why VS Code on That List Changes Everything

Here's what I keep coming back to. Gmail and GChat are one thing — capturing how people write emails and schedule meetings is fairly pedestrian training data. But VS Code?

They're recording how developers write code. Not just the output, but the micro-behaviors that define how an experienced engineer works: the way they grep through a codebase, the order they check files when debugging, the refactoring patterns they reach for instinctively, how they bounce between terminal and editor. All of that is now training data for computer-use agents.

This is a fundamentally different animal from the Copilot controversy. GitHub scraped public repositories — code that was already visible, even if the license implications were murky. What the MCI captures is private workflow patterns from employees who have no opt-out. The model isn't learning from what you wrote. It's learning from how you work.

And honestly, the approach makes cold technical sense. Computer-use AI has been a breakthrough area this year. GPT-5.4 scores 75% on OSWorld, beating the human baseline of 72.4%. Claude models push even higher — Opus 4.7 hits 78% on the verified subset. But these benchmarks run in controlled environments: clean desktops, predictable application states, synthetic task descriptions. Real-world computer use is messier. Partially loaded pages, unexpected modal dialogs, muscle-memory shortcuts no documentation captures, the weird tab-switching pattern you do when you're stuck on a bug.

Getting from benchmark-superhuman to production-reliable requires exactly the kind of naturalistic data that thousands of employees generate every workday. The uncomfortable part isn't that the approach is wrong. It's that it works.

The Timing Is Not Subtle

Meta is simultaneously committing $135 billion in AI capital expenditure for 2026 and reportedly planning to cut 20% of its workforce starting in May. Bloomberg's headline nailed it: "Meta Is Making Workers Train Their AI Replacements."

The company acquired a 49% stake in Scale AI for $14 billion last year, and Scale's former CEO Alexandr Wang now runs Meta Superintelligence Labs. Zuckerberg has been explicit that AI should subsume an increasing share of engineering work at the company. The MCI isn't a research curiosity — it's infrastructure for that transition.

And this won't stay contained to one campus. The entire premise of computer-use AI — the capability that OpenAI, Anthropic, Google, and now Zuckerberg's team are all racing toward — demands massive datasets of real human-computer interaction. Public benchmarks got the industry to 75% task accuracy. Getting to 99% probably requires what the MCI does: recording how real workers perform real tasks in production, at scale, across thousands of different behavioral patterns.

The program only targets U.S. employees because European privacy law makes this legally radioactive. Italy bans electronic employee tracking for productivity purposes outright. But in most American jurisdictions, if it's a company-owned device and the employer has disclosed monitoring, almost anything goes. The legal guardrails that would normally constrain this kind of data collection simply don't exist yet for the "training AI on employee behavior" use case.

What to Watch

The question for every developer isn't whether this practice spreads — it's whether your employer will be transparent when it does. At least the Menlo Park team published a memo. If you work at a company with AI ambitions and haven't checked your monitoring disclosures lately, this would be a good week to pull up that acceptable use policy.

The agents being trained on your colleagues' workflows today will ship as products tomorrow. That's the business model.