Meta just did something quietly dramatic: it stopped grading its employees on how much they use AI. The same week, workers started testing the Meta Hatch agent, a new AI tool that can browse the web and use apps on its own. These two moves tell you everything about where AI at work is heading, and neither one is what the headlines make it look like.
What is Meta’s Hatch agent?
Hatch is Meta’s internal project for a personal AI agent. It’s the kind of tool that’s been getting a lot of attention lately: you give it a task, and it autonomously carries out actions on your computer. It browses websites, clicks through applications, and gets things done without you babysitting each step. The Meta Hatch agent is, in that sense, the most direct example yet of an agent built for regular office work rather than developers.
If that sounds familiar, it’s because Hatch is essentially Meta’s version of OpenClaw, the open-source agent that went viral earlier this year. Meta employees have been testing Hatch on their corporate devices for weeks, and Wired reported an expected public release. So this isn’t just an internal experiment anymore. It’s a product coming to you.
What kinds of things can it do? Based on what’s been reported, think of tasks like booking appointments, organizing your calendar, or handling multi-step web tasks that would normally take you twenty clicks and a spreadsheet. Some employees have already embraced it for exactly that kind of personal organization.
Why Meta stopped ‘tokenmaxxing’
The other half of this story is the end of a strange experiment. About a year ago, Meta told employees they’d be evaluated on “AI-driven impact.” In practice, that meant your performance review could depend on how much you used AI tools, with labels floating around like “AI Native,” “AI First,” and “AI Enabled.”
Employees responded the way humans do when you attach a score to behavior: they game the system. People prompted AI tools as much as they could, sometimes frivolously, just to pump up their usage numbers. One worker built an internal leaderboard tracking AI usage, complete with titles like “Token Legend.” It leaked, went viral internally, and got taken down in April.
This week, Meta formally ended it. New guidance says outcomes “can be supported by AI or other means,” and engineers were told the company “will not use AI adoption dashboards or token counts to evaluate impact.” In other words, grading people on token consumption was a failed experiment, and Meta is quietly admitting it.
The context matters here. A lawsuit filed in July by about two dozen former employees argues that Meta’s focus on AI usage violated antidiscrimination law, because people on health or family leave couldn’t accumulate usage and got penalized in layoffs. Meta denies the allegations. The timing of this policy change, right after that lawsuit and after months of “tokenminimizing” to cut AI costs that reached billions, is probably not a coincidence. What you’re watching is a company learning, in public and expensively, that a Meta Hatch agent or any other tool only helps when people actually want to use it.
What Hatch can actually do
Let’s get concrete about the agent itself. From the reporting, the Meta Hatch agent can:
- Browse the web and operate other applications
- Handle multi-step tasks that require moving between tools
- Work on your behalf while you do something else
The catch is the same one every personal agent faces: it needs access to your accounts to be useful. And that’s where Meta’s employees got hesitant. Some said they don’t want to connect their personal email and calendar to Hatch, worried about privacy and about the AI messing up something important.
That hesitation is worth remembering. Meta has history here. The company previously ran a project that tracked employees’ keystrokes and other activity on work devices to collect data for training AI systems. It paused that project after an internal security breach, but the trust damage stuck. Employees who lived through that aren’t exactly eager to hand a new agent the keys to their inbox.
What it means for your job
Step back and look at the big picture. In less than a year, Meta went from forcing AI usage (and grading it) to rationing it, to this week’s backpedal on token counting, all while rolling out an agent that fits squarely into the trend of persistent AI coworkers that we’ve been tracking.
The takeaway for regular workers is more interesting than the drama. If a company as big as Meta can’t make “use more AI” a performance metric without breaking, your employer probably can’t either. The meaningful question isn’t whether you use AI, it’s whether you use it well. That’s the signal this week’s change sends.
But there’s a harder truth underneath. The Reuters report from last week said Meta was gearing up for another round of major layoffs, on top of the 8,000 jobs cut in May, and only called it off because of a spike in software bugs linked to heavy AI reliance. Zuckerberg says no further mass layoffs are expected this year, but the direction is clear: agents like Hatch that boost productivity are exactly the kind of thing that makes companies reconsider headcount, even when they don’t cut.
If you want to stay ahead of that curve, the practical move is to learn how personal agents actually work before they arrive in your workplace. Messing around with OpenClaw or any other agent on your own machine is the best insurance you can buy (and yes, you can run agents safely), because the people who understand these tools are the ones who end up directing them instead of being directed. Grok Bot, which we’ve covered before, is another example of the same idea: an agent platform you can start using today without waiting for a Meta rollout.
Bottom line
The Meta Hatch agent is coming, and it will bring the same questions it raised internally: what do we let it touch, and what does it do to jobs? The company’s decision to stop grading AI usage is a rare moment of honesty from a tech giant: forcing people to use AI doesn’t work.
Use the breathing room to get comfortable with agents on your own terms. By the time your workplace rolls one out, you’ll already know what it’s good for, what it shouldn’t touch, and how to make it prove its worth before you trust it with anything important.
