Amazon tells engineers to stop gaming AI tool usage and focus on what software they ship

In short

Amazon has revised its engineering standards to advise teams on evaluating AI coding tools based on the quality of software produced rather than the volume of tokens used. Tokenmaxxing means deliberately using more AI tool requests to look productive on internal dashboards rather than actually improving work.

• Why it matters: Companies are spending huge amounts on AI tools but seeing little useful output in return. Amazon's shift means engineers will be judged on real results, not on appearing busy with AI.

What tokenmaxxing is

Earlier this year, Amazon employees were using an internal AI agent called MeshClaw to increase their token counts on company dashboards. MeshClaw can deploy code and interact with Slack, the messaging platform. Employees were essentially gaming the system by running AI tasks just to boost their numbers ¹.

Amazon responded by shutting down an internal leaderboard in May after staff gamed the rankings. The company has not imposed hard caps on employee token usage ¹.

How Amazon wants to measure success

The new guidelines say Amazon should measure long-term outcomes instead of chasing raw usage numbers. Managers are asked to look at token consumption alongside other indicators, not as the main target ¹.

Amazon recommends examining teams at both extremes. Low AI usage might mean engineers are missing chances to improve their work. Heavy usage should raise a different question: is the delivery actually improving enough to justify the cost? Neither number alone tells the full story ¹.

The guidance also makes clear that deciding a task does not benefit from AI tools is "a finding, not a failing." Explaining why a tool has limits is seen as showing judgment, not resisting change ¹.

Other companies face the same problem

Amazon is not alone in seeing a gap between AI spending and useful results. Uber disclosed in May that it had used up its entire 2026 budget for Claude Code, Anthropic's AI coding assistant, by April — and president and COO Andrew Macdonald admitted that higher token usage had not yet produced useful consumer features ¹.

In June, leaked audio obtained by 404 Media showed Accenture executives discussing rising token expenses as clients shifted from basic chatbots toward more advanced AI systems capable of independent action, according to TechSpot. The executives said they had doubts about two things: the price of running individual tasks and whether the money being spent was actually worth it. The company was reportedly working on an advisory tool named Token IQ ¹ aimed at helping clients better control those expenses.

Amazon reaffirmed that AI will remain central to software engineering practice, whatever best practices look like in the future ¹.

Words to know

Token: A unit of text that AI models process; one token is roughly four characters of text · Tokenmaxxing: Using AI tools excessively just to appear productive on company dashboards ¹

Key facts

• Amazon updated engineering guidelines to judge AI tools by delivered software, not token usage
• Employees previously used MeshClaw AI agent to game internal token dashboards
• Uber exhausted its entire 2026 Claude Code budget by April with no useful features to show
• Amazon shut down an internal AI leaderboard in May after employees gamed the rankings
• Accenture is developing Token IQ tool to help clients manage rising AI spending

Sources

1. Amazon tells engineers to stop tokenmaxxing and judge AI tools by what they ship — techspot.com