Updated guidance tells Amazon managers to judge AI-assisted engineering by the quality of software delivered, after employees reportedly gamed usage leaderboards.
Amazon is telling its engineers to treat AI coding tools as a means to better software, not as a metric to be chased for its own sake. In updated engineering guidance, the company says teams should be judged on the quality of the systems they ship, not on how many tokens they burn through, how many code changes they generate or how often they rely on AI-assisted reviews. The message is clear: internal adoption data matters, but only when it is read alongside delivery outcomes and engineering judgement.
The shift is aimed at a familiar workplace distortion: “tokenmaxxing”, where employees inflate AI usage to make themselves look more productive on dashboards. According to TechSpot, Amazon has not introduced token caps, but it has warned managers not to treat usage levels as a standalone performance signal. Very low use may indicate that engineers are missing opportunities to improve their work, while very high use raises a different question altogether: whether the added consumption is actually helping the product move faster or better. The company also says that deciding a task is not suited to current AI tools should be seen as valid evidence, not as a failing.
The guidance follows reports that Amazon employees were exploiting internal tools to boost their token totals. Tom’s Hardware reported that staff used MeshClaw, an internal AI agent that can deploy code and interact with Slack, to push up usage figures, prompting Amazon to shut down an internal leaderboard in May. Livemint separately reported that the leaderboard was tied to the company’s Kiro agentic development platform and had been removed after users gamed it with unnecessary agents. Amazon has also said, in the guidance quoted by TechSpot, that “we believe that it is best to measure, and be measured by, long-term outcomes”, while adding that AI will remain central to software engineering practice.
The episode reflects a broader problem across the technology industry: token consumption does not always map cleanly to productive output. Tom’s Hardware reported that similar concerns have surfaced at Meta and Microsoft, while TechRadar noted that aggressive usage targets and leaderboards can create perverse incentives without guaranteeing meaningful business value. More broadly, companies are still working out how to price and monitor agentic AI systems as they move from simple chat interfaces to code-writing and workflow automation, a transition that can drive up costs before it delivers clear gains.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
