Amazon employees 'tokenmaxxing' MeshClaw use as adoption targets and leaderboards pressure behavior

Workers are widely deploying an internal AI agent platform called MeshClaw and sometimes automating non‑essential activity to inflate measured AI token consumption — behavior dubbed “tokenmaxxing”—driven by internal targets and visible usage statistics.

Amazon employees have increasingly used an internal agent platform, MeshClaw, to automate tasks and to boost measured AI activity, a practice some colleagues call “tokenmaxxing.” Three people familiar with the matter say staff deploy agents not only to speed work but also to generate extra token consumption as a signal of AI use to managers, raising concerns about how internal metrics shape behavior.

MeshClaw was built by more than three dozen Amazonians and allows users to create agents that connect to workplace software and act on a user’s behalf. People familiar with the tool say it can initiate code deployments, triage email, and interact with apps such as Slack. The product was inspired by OpenClaw, and internal memos describe MeshClaw as a way to consolidate learning and monitor deployments overnight. Employees described two overlapping uses: practical automation that replaces repetitive tasks, and deliberate activity intended primarily to raise token counts. Some staff admitted automating non‑essential actions specifically to increase usage statistics; colleagues have adopted the slang “tokenmaxxing” to describe that deliberate inflation of token consumption.

The push to show AI adoption has official roots. Amazon set targets that aim for more than 80% of developers to use AI each week and introduced internal tracking of AI token consumption with leaderboards. That emphasis mirrors broader Silicon Valley efforts to demonstrate returns on big infrastructure investments, and employees at other firms have reported similar token‑inflating behavior on internal metrics.

Those visible metrics and the social pressure they create appear to have produced perverse incentives. Employees said managers do look at usage statistics even though Amazon has told staff that token numbers would not be used in performance reviews. Team‑wide stats were once visible across groups and were later restricted so only individuals and their managers can see their own numbers, but several current employees said they still believe monitoring affects day‑to‑day behavior.

Staff also raised concrete security concerns about agents granted permission to act autonomously. Multiple employees warned that the default security posture and the potential for unintended actions or errors are worrying; one current employee said they were not comfortable letting an agent “just do its own thing.” Those worries reflect both the operational risks of automated actions and unease about how broadly permissions are granted.

Amazon said MeshClaw enables thousands of employees to automate repetitive tasks daily and framed the tool as part of responsible AI deployment, while discouraging managers from using token counts to assess performance. The company’s response highlights a tension between promoting automation at scale and preventing metrics‑driven behaviors that can distort work priorities.