Meta drops AI usage metric and rolls out Hatch autonomous agent for staff

Meta announced this week that performance assessments will no longer be based on the volume of AI tool usage, and at the same time began allowing employees to test a new autonomous agent capable of browsing the web and launching applications. The move ends a period in which workers were measured by the number of tokens they consumed, a phenomenon dubbed “Tokenmaxing,” and shifts the focus back to actual impact regardless of the underlying technology.

About a year ago Meta said employees would be evaluated according to “AI-driven impact,” which in practice meant measuring the extent of chatbot and agent usage. Workers were tagged as “AI Native,” “AI First” or “AI Enabled.” According to a claim filed by about twenty-four employees in July, those on health or family leave could not accrue usage and were harmed in their evaluations, claims the company denies. New guidelines released this week replace those criteria with a vague wording that “these results may be supported by AI or other means,” returning the emphasis to output itself.

In recent months employees documented a pattern of repeated instructions for tools, sometimes appearing artificial, solely to inflate internal consumption metrics. The peak of the practice occurred when one employee created an internal leaderboard naming the top users “Token Legend.” After the details leaked, the board was removed in April, and a few months later the company began limiting AI-tool usage quotas. Company spokesperson Tracy Clayton said the updates were meant to emphasize what has always been policy, evaluation by contribution, and noted that the labels never served for performance assessment. According to a report by The Information, engineers were told the company “will not use AI adoption dashboards or token counts to assess impact.”

At the same time Meta is promoting, but not mandating, use of its most advanced experiment to date: an agentic tool called Hatch, reminiscent of the viral project OpenClaw. The tool can browse the internet and run other applications on a computer, and employees have been testing it on company devices for several weeks ahead of a public launch expected later this year. Some staff are hesitant to connect Hatch to email, calendar and personal services over privacy concerns and the risk that the agent might take erroneous actions with consequences. A previous project that tracked keystrokes and activity on work devices for training data—frozen since—already damaged employee trust in the company’s AI tools.

Removing the threat of a negative evaluation for avoiding AI checks may improve morale and enable a rebuilding of trust. Already a number of employees report using Hatch to schedule personal meetings and organize private life, describing the tool as efficient.