Imagine you’re at a job where your boss once told you the yardstick of success was how many rounds you could fire. You emptied the magazine until supply chains hiccuped and the company quietly circulated a memo saying, actually, don’t shoot so much. Then someone hands you the minigun from Terminator 2 and smiles.
I’ve been watching this at Meta like a slow-motion office sitcom with real consequences. You and I both know companies send mixed signals all the time, but this one smells like strategy and panic in equal measure.
I watched an engineer open Hatch at their workstation: Meta is both pushing AI and pulling back on “tokenmaxxing”
Here’s the plain scene: internal dashboards once tallied token counts like sales figures. Leaders celebrated heavy AI use. Then reports from Kevin Roose at the New York Times and pieces in Wired and The Information suggested something changed—the scoreboard was quietly being taken down.
I’ve talked to people who felt pressure to hit token quotas, and others who breathed easier when those quotas vanished. That tug-of-war explains why you see memos saying “use AI where helpful” while product teams quietly hand engineers access to a tool that guzzles tokens.
What is tokenmaxxing?
Tokenmaxxing was the shorthand for treating API tokens as a KPI: the more you burned, the more visible and valued you became. OpenClaw-style agent platforms and internal leaderboards turned consumption into a sport. At Meta, Wired reported that performance guidance now says outcomes “can be supported by AI or other means,” which feels like permission to choose—if you can trust management not to change the rules again.
A colleague demoed Hatch in a meeting: It’s an agentic tool that behaves like a greedy app
Observation: someone slid their screen around the table and showed Hatch running multiple agents at once.
Hatch, as reported by The Information and Wired, looks like Meta’s answer to agent platforms such as OpenClaw. It consumes more tokens than a chat interface or a code assistant. That makes it powerful—and expensive. It also feeds the central tension: if your internal tool saves time but inflates token metrics, what does management reward?
I read the Reuters report aloud at lunch: The layoffs rumor hangs over every AI pilot
Observation: a group at the table went quiet when Reuters reported plans to cut up to 60% of staff in some teams if AI could replace them.
That scare is still raw. Reuters later said the plan stalled because the software wasn’t reliable enough to take over jobs. Still, when a tool like Hatch is used internally, people imagine their roles being replaced. You don’t need to be a pessimist to worry that internal pilots can become an argument for headcount reduction.
Will Meta replace employees with AI?
Short answer: not wholesale, at least not yet. Meta’s internal experiments have bugs, and leadership knows replacing humans with agents is messy. But the specter is real: when an AI tool can perform parts of a job and costs less than salaries, it becomes a tempting lever. That’s why engineers who favor Hatch—Alexandr Wang’s memo about Slack being superior for agents is part of this story—are also nervous when keystroke-tracking experiments leak into public view.
I watched a Slack channel switch from Google Chat: Tools shift to favor agents
Observation: an internal memo explained why Slack replaced Google Chat—because it’s friendlier for agents.
Switching platforms is how a company primes itself to scale agent usage. Slack presents conversation threads and integrations that make agent orchestration feel natural. If you’re an employee, the tools you use sculpt your day: they nudge you toward or away from scripted, repeatable workflows that agents can eat.
A reporter at TechCrunch got pinged by Instagram: People notice privacy problems quickly
Observation: Amanda Siberling used the Meta AI web app for a story and Instagram notified her friends.
Privacy stumbles make people recoil. When internal programs considered using mouse and keystroke data to train models, employees who already feared being judged for token counts suddenly feared being mined for models. Meta paused that program after scrutiny, but the damage to trust can linger.
Two metaphors will help you hold the contradiction in your head: this is like handing someone a minigun while asking them to conserve ammo, and like a thermostat that flips between freezing and boiling as leadership changes its mind. Those images capture the cognitive whiplash employees report.
For search intent: this story touches on OpenClaw-style agents, Hatch, Meta AI, Slack, Google Chat, OpenAI, Mark Zuckerberg, Alexandr Wang, Kevin Roose at the New York Times, Wired, The Information, Reuters, Business Insider, and TechCrunch. If you’re trying to figure out whether to trust internal AI tools, watch three signals: whether token counts reappear in performance metrics, whether tools are rolling out company-wide, and whether privacy experiments are transparently explained.
Meta’s message is currently mixed: encouragement without mandate, secretive pilots and public pullbacks. You can see why employees oscillate between enthusiasm and fear, and why investors and regulators are watching the playbook unfold—will the company promote agent productivity as a path to layoffs, or will imperfect AI simply augment people’s work?
What would you do if your company handed you a minigun and told you to be careful with the ammo?