Meta Pushes Its New AI Agent on Employees—but Eases Off on Tokenmaxxing

Paresh Dave and Maxwell Zeff report Meta told workers this week that performance evaluations will no longer depend on how much they used AI tools, according to three employees. New review language replaces 'usage of AI' and 'AI Native' labels with a caveat that outcomes 'can be supported by AI or other means.' Spokesperson Tracy Clayton says Meta always evaluated contributions and that 'AI Native' was never used for performance. The Information reported engineers were told dashboards and token counts will not evaluate impact. A July lawsuit by about two dozen employees alleges Meta violated antidiscrimination law in May layoffs, saying workers on health and family leave could not accumulate usage; Meta denies the allegations.

For months, some staff prompted tools frivolously to max internal token counts. One worker built a leaderboard with labels like 'Token Legend,' took it down in April after it leaked, and Meta later rationed usage. Employees are now testing Hatch, an agentic tool similar to OpenClaw that can browse and operate apps, on corporate devices ahead of an expected public release. Some hesitate to connect personal email and calendars, citing privacy and a paused project that tracked keystrokes on work devices for training data. Hatch still burns more tokens than chatbots. Reuters reported Meta was preparing more layoffs after 8,000 cuts in May, then pulled back amid AI-linked bugs and slow agent progress. Zuckerberg has said further mass layoffs are not expected this year.

The contradiction is the story. Tokenmaxxing as a performance sport is over, at least on paper, because it was stupid and because it showed up in a discrimination complaint. Hatch is the replacement pressure: not a dashboard grade, an agent on the laptop that management still wants people to 'wield effectively.' Clayton's line that labels were never for evaluation sits poorly next to a lawsuit that says they were, in practice.

Hatch on personal Gmail is the trust test Meta already failed once with keystroke logging. Employees booking their own appointments with a company agent is a data-flow decision disguised as productivity. The ones who refuse are the ones who remember the last instrumentation project.

Signal Desk's read: Meta is trying to keep the AI-native culture after proving that measuring it by tokens produces junk prompts and legal risk. Easing the grade is rational. Shipping Hatch internally before a public release is still a forced march. If the agent works, some staff expect it to justify the next cut — Reuters had the round planned, then bugs intervened. Zuckerberg's 'not this year' is a calendar, not a promise about 2027.

The Token Legend leaderboard should be taught in management school as how not to instrument a workforce. That Meta needed a year to stop grading people on a meter is the cultural tell. Hatch will create a new meter whether or not reviews cite it.

Almost a year ago Meta told workers they would be graded on AI-driven impact, which in practice meant chatbot and agent usage. The backlash, the leaked leaderboard, rationing, a layoff-discrimination suit, and now a quieter rubric are one arc. Hatch is the product that arc is supposed to birth.