Meta tracks workers for agent data

Meta is treating its own workforce as a behavioral data source for agent training, which is technically logical and culturally radioactive.

• –Real computer-use agents need more than screenshots; they need dense traces of how people navigate menus, shortcuts, forms, and messy enterprise workflows.
• –The privacy concern is not theoretical: keystrokes and screenshots can capture credentials, source code, customer data, and workplace communications unless filtering is extremely strong.
• –This is a preview of where agent training is heading: companies with large employee bases and internal tools can build proprietary UI-action datasets outsiders cannot easily replicate.
• –Developers should expect sharper internal-data governance fights as AI teams push beyond text logs into full interaction telemetry.