Staff Reports
Amazon quietly posted updated guidance for its software engineers this week telling them to stop gaming AI usage numbers — the practice the kids on the internet call “tokenmaxxing.” The memo says raw token counts, flashy adoption rates and bloated code-throughput numbers are “deeply flawed” ways to measure engineering work. In plain English: stop burning compute just to look busy.
Amazon tells engineers: stop “tokenmaxxing”
The new role guideline, circulated internally before it went public, calls out tokenmaxxing by name and ties the problem to Amazon leadership principles like frugality and judgment. It even points to an embarrassing episode earlier this year when an internal leaderboard encouraged people to pump up their AI token use. That leaderboard is gone now, and the memo bluntly tells managers to focus on long-term outcomes instead of surface metrics. If your code tasks don’t benefit from current AI tooling, the document says, that is a finding — not a failing.
Why this memo matters for engineers and taxpayers
Metric design drives behavior. Give people a leaderboard and some will chase it with bots, agents, or expensive model calls that add zero customer value. The costs are real: cloud compute bills climb fast when teams let fancy models run wild. Amazon’s AWS leadership warns they judge AI spending on return on investment and “feature velocity” — how many real features ship and how many successful changes reach customers. Translation: productivity looks like shipped work, not token consumption or selfie-style AI demos.
Balance experimentation with discipline
We should want teams to experiment. Innovation needs room to fail fast and learn. But “fail fast” shouldn’t mean “bill us later.” Good policy is simple: tie incentives to real customer outcomes, add guardrails for runaway agents and background jobs, and stop using flashy dashboards as hidden promotion signals. Managers should be held accountable for explaining both successes and the experiments that failed without delivering value. That’s common sense — and, yes, it’s conservative sense too: get more out of less.
Takeaway: call out waste, preserve innovation
Amazon’s memo is an encouraging move. It slaps down a game that made some engineers look productive while wasting company resources. But a memo isn’t a cure-all. Companies need clearer governance, transparent measurement, and honest promotion systems so tokenmaxxing never sneaks back in. If Silicon Valley wants to sell us on AI as a productivity miracle, it must first stop treating token counts like participation trophies. Frugality and accountability are not buzzwords; they’re how value is actually created.
