The agents aren't just doing the work — they're learning by watching you do it first.
The Summary
- Meta is installing tracking software on U.S. employee computers to capture mouse movements, clicks, keystrokes, and screen snapshots — all to train AI agents that can perform work tasks autonomously
- The tool runs on work apps and websites, part of Meta's "Model Capability Initiative" under its SuperIntelligence Labs
- This drops weeks before a planned 10% global workforce reduction starting May 20, with more cuts expected later in 2026
- The real tell: we don't know how far up the org chart the surveillance goes
The Signal
Meta just made the quiet part loud. The Model Capability Initiative isn't about productivity monitoring in the traditional HR sense. It's about harvesting the ground truth of how knowledge workers actually work, so AI agents can replace them.
This is training data collection at scale, and it's happening on corporate hardware where consent is functionally coerced. Every click is a labeled example. Every keystroke sequence is a workflow annotation. Every screen snapshot captures context that's impossible to get from clean datasets.
"The agents aren't learning from job descriptions. They're learning from watching you navigate the actual chaos of your day."
Compare this to how previous automation waves worked:
- Factory automation required engineers to map physical processes
- Software automation required analysts to document workflows
- AI automation just watches you work and learns the pattern
The efficiency gain isn't just technical. It's methodological. Meta doesn't need to pay consultants to document how their employees work. They're keylogging the documentation in real time.
The timing matters. Meta plans to cut 10% of its workforce starting May 20, with more reductions coming. That's not coincidence. You don't collect this much behavioral data unless you're planning to automate the behaviors. The severance packages and the training data extraction are two parts of the same transition plan.
The Implication
If you work at a company building AI agents, assume your work is training your replacement. The knowledge work that's most legible to capture — repetitive, screen-based, rule-following — is the knowledge work most at risk. The question isn't whether your employer will do this. It's whether they'll tell you when they start.
For everyone else: watch which companies follow Meta's lead. Enterprise software vendors will productize this within 18 months. The playbook is written. Capture interaction data. Train task-specific agents. Reduce headcount. The companies that move fastest will have the competitive edge. The workers who adapt to agent-augmented roles will be the ones who stay.