Big quote: Amazon is telling its engineering teams to judge AI coding tools by the software they help deliver, not the volume of tokens they consume. The updated engineering guidelines warn against treating AI adoption, token consumption, code changes and code reviews as performance targets. "We believe that it is best to measure, and be measured by, long-term outcomes," Amazon wrote.
Amazon is not abandoning usage data. It is asking managers to examine those figures alongside other indicators.
The company recommends examining teams at both ends of AI adoption. Low usage may suggest engineers are missing opportunities to improve their work. Heavy usage should prompt a different question: whether delivery is improving enough to justify it. Neither figure settles the question on its own.
The guidance targets tokenmaxxing – deliberately inflating AI consumption to appear productive on internal dashboards.
Earlier this year, Amazon employees were using MeshClaw, an internal AI agent that can deploy code and interact with Slack, to increase their token counts. Amazon shut down an internal leaderboard in May after employees gamed the rankings. Amazon says it has not imposed employee token caps.
Amazon's guidance leaves room for engineers to conclude that a particular task does not benefit from the available tools. It describes that conclusion as "a finding, not a failing," treating an explanation of a tool's limitations as evidence of judgment rather than resistance to adoption.
"We are certain that whatever future best practices are, AI will be at the center of software engineering practice," the company wrote.
Other companies are confronting similar gaps between AI consumption and useful output. Uber president and COO Andrew Macdonald said in May that higher token usage had not yet translated into useful consumer features. His comments followed the company's disclosure that it had exhausted its 2026 Claude Code budget by April.
At Accenture, leaked audio obtained by 404 Media in June captured executives discussing rising token spending as clients moved from chatbots to agentic workflows. Executives discussed uncertainty about both the cost of individual tasks and whether the spending was worthwhile. The firm was reportedly developing an advisory tool called Token IQ to help clients manage those expenses.
Amazon tells engineers to stop tokenmaxxing and judge AI tools by what they ship
