Amazon tells engineers to stop 'tokenmaxxing' to look productive

Amazon’s updated engineering guidelines reject “tokenmaxxing”: deliberately using more AI or generating more tokens to appear productive on internal dashboards. The company says adoption rates, token counts, code changes and review volumes can mislead when treated as goals rather than considered alongside actual results. Such targets may encourage wasteful tool use, increase computing costs and make software more complex. Amazon instead says performance should be judged by long-term outcomes, while usage data can still offer context when interpreted carefully. Managers are encouraged to compare teams with high and low AI use to see whether adoption improves delivery.

The guidance does not signal a retreat from AI: Amazon expects it to become central to software engineering, but says not every task needs an AI tool. Finding that current tools do not help a particular job is a valid assessment, not a failure to adopt. The company says it has not set employee token limits, though it shut down an internal leaderboard after workers reportedly gamed it, including by using an AI agent to inflate counts. The issue reflects wider concerns that rising AI consumption does not necessarily produce better work: Uber has questioned the link between token use and useful features, while consulting firms and AI companies are grappling with unpredictable, escalating costs.