Product1 distinct publisher3 min readPublished
Meta has taken AI adoption dashboards out of engineer evaluations, according to The Information, at the same time as Wired reports staff are testing Hatch, an internal agent that burns more tokens than a chatbot does.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
An engineer who spent the spring climbing a leaderboard now finds that leaderboard deleted from the review form. In March, according to Kevin Roose's New York Times column, Meta was rewarding workers who made heavy use of AI tools and chastening those who did not, with leaderboards turning usage into a contest [3]. The guidance Wired describes now says desired outcomes "can be supported by AI or other means" [1]. That wording lets Meta drop the number without saying the number was wrong.
Token counts get used because they are cheap to collect and easy to rank. What they measure is volume. What they leave out is whether the work landed, or how far into a real workflow the tool actually got, or who would still open the thing in a quarter when nobody is counting, which is the only question a finance team eventually asks.
The constraint that moved was not the metric. The Financial Times reported in June that Google was throttling Meta's AI tokens, forcing the company to rethink tokenmaxxing [4], roughly three months after the leaderboards became public knowledge [12]. Meta could retire its own metric with a document. It could not do the same to a supplier's rate limit, which had to be negotiated instead.
Separate the thing being pitched from the thing being done. The pitch is that usage is no longer assessed [2]. The practice is that employees have had Hatch for several weeks, encouraged but not required, and Wired says it consumes more tokens than standard chatbots and AI coding tools [5]. The Information reported in May that Hatch is Meta's own version of an OpenClaw-style agentic platform [6]. Spend moves off the person typing and into a process that keeps running after they stop.
"Encouraging but not requiring" does different work depending on what else is in the room. AI chief Alexandr Wang's memo this week explained the move from Google Chat to Slack as a way to enable agents, calling Slack "the strongest platform available today for agents" and telling staff that everyone at the company will benefit from an agent ecosystem even though not all of them are building agents today [8]. Wired's sources say the fear is that internal success for Hatch is followed by layoffs [9], and Reuters reported last week that Meta had planned to cut up to 60% of the workforce in certain areas, a plan that stalled on bugs in the software meant to take over [10]. Some staff are also wary because of April's disclosed plan to train AI on employee keystrokes, which Meta later paused [7]. Others told Wired they are simply relieved not to feel pressure to use AI constantly [11]. Meta has not put its own account on the record here; Gizmodo says it contacted the company [13].
For anyone rolling out a mandated tool on Monday, the forcing function is a two-by-two. One axis: whether usage sits in someone's review today. The other: whether that person would keep using the tool if the count disappeared tomorrow. Counted-and-would-keep is adoption you can stop measuring. Counted-and-would-stop is compliance, paid for twice, once in tokens and once in the goodwill it consumed. Uncounted-and-still-using is your evidence base. Uncounted-and-not-using is your honest baseline. Meta has just moved a large engineering organisation into the bottom row deliberately, and how much Hatch usage survives that move is the only adoption figure from this episode worth quoting to anyone.
Ranked by verification strength, evidence, and original report placement.
Wired, citing anonymous Meta insiders, reports that newly published performance review guidance for Meta employees no longer makes "usage of AI" a criterion for assessment, and says desired outcomes "can be supported by AI or other means".
The Information reported Wednesday that Meta engineers had received word that "AI adoption dashboards" and "token counts" would no longer be part of evaluations.
A March column by Kevin Roose in the New York Times said Meta was "rewarding workers who make heavy use of A.I. tools and chastening those who don't", and mentioned leaderboards that turned AI usage into a competition; the practice was dubbed tokenmaxxing.
A June report from the Financial Times said Google was throttling Meta's AI tokens, forcing the company to rethink tokenmaxxing to some extent.
According to Wired, Meta has been encouraging but not requiring employees to use Hatch, its most advanced AI experiment; Hatch is still unreleased, employees have had access for the past several weeks, and it consumes more AI tokens than standard chatbots and AI coding tools.
In May, The Information reported that Meta was developing Hatch, a consumer-friendly version of an OpenClaw-like agentic AI platform.
Distinct publishers with included, body-backed reporting in this cluster.
gizmodo.com
1 article · September 3, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
product
Meta scraps the AI usage test from its performance reviews1 distinct publisher
product
Meta's Mac app is a data connector wearing a chatbot's clothes6 distinct publishers
invest
Meta's Hatch agent tops out at $199.99 a month, with DoorDash and Etsy behind the meter2 distinct publishers
product
Meta budgeted up to $10bn a year with Anthropic even as Zuckerberg warns against rival AI labs1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One relay, unnamed insiders
Nothing in this story was reported to us first-hand. Gizmodo is aggregating: Wired's anonymous Meta insiders carry the rewritten review guidance and the Hatch rollout, The Information carries the dashboards-and-token-counts line, and the Financial Times, New York Times and Reuters details arrive as citations. The one thing Gizmodo did itself was ask Meta, and Meta said nothing. The quoted guidance language is the strongest artefact on offer, and even it reaches us paraphrased.
Real but wholly internal
There is genuine movement, all of it behind Meta's badge readers: an unreleased agent in employees' hands for several weeks, a company-wide chat migration justified as agent infrastructure, and an evaluation rubric already edited. What is absent is any measure — no headcount on Hatch, no token volumes, no ship date, and no sign the consumer product this was meant to become exists yet.
Drama ahead of confirmation
Gizmodo's own framing is dry — mixed messages, not a revolution — and it takes care to note Zuckerberg's unmet billion-user promise. The overstatement sits in the load the story asks unnamed sources to bear: a 60% cut plan in certain areas, shelved over software bugs, and layoffs said to follow if Hatch works internally. Those are the sentences readers will remember, and they are the least verified things in the piece.
Token bills and job security
Follow the cost. Google metered Meta's tokens in June; by September the metric that encouraged staff to burn them is gone, while a hungrier agent is being handed out voluntarily rather than mandated — which happens to shift the consumption from a scored duty to a personal choice. The employees talking to Wired have their own stake on both sides: relief at losing the mandate, and fear of being automated by the thing they are helping test. Meta's silence leaves every reading of its motives to other people.
Plausible, thinly held
The core of it hangs together — two outlets describe the same reversal, the throttling supplies a motive, and the timeline runs in the right order. But the whole picture comes through a single retelling of other people's sources, with the most consequential numbers unverified and the company mute. Enough to act on directionally; not enough to quote as settled.