Amazon employees automate tasks with MeshClaw amid AI usage pressure

Amazon set a target: more than 80% of its developers should be using AI tools on a weekly basis. To track progress, the company built token consumption leaderboards. And then, as anyone who has ever watched a metric become a target could have predicted, people started gaming the system.

The company’s internal AI tool, called MeshClaw, has been widely deployed in recent weeks, allowing employees to spin up AI agents that connect to workplace software and execute tasks on their behalf. Thousands of employees now use it daily, according to Amazon. The problem is that a meaningful chunk of that usage appears to be performative rather than productive.

MeshClaw lets users create agents that plug into tools like Slack, handle email triage, and manage code deployments. In practice, employees have started automating non-essential tasks just to show managers they’re engaging with AI more frequently.

The phenomenon has earned an internal nickname: “tokenmaxxing.” In English: burning through AI tokens on trivial work to climb the usage leaderboards that managers reportedly monitor.

One anonymous employee described “so much pressure” and “perverse incentives” driving the behavior.

Amazon has said the usage statistics won’t factor into performance evaluations. But telling employees their metrics don’t matter while simultaneously displaying those metrics on leaderboards that managers can see is a bit like putting a plate of cookies on the table and saying “those aren’t for you.”

Amazon’s push isn’t happening in a vacuum. Every major tech company is racing to embed AI into its enterprise workflows, following patterns seen in Google’s internal agent tools and Microsoft’s Copilot expansions.

For Amazon specifically, demonstrating internal AI adoption serves a dual purpose. It makes the case that AI is transforming productivity inside the company, and it creates a proving ground for AI tools that could eventually be packaged and sold to AWS customers.

The crypto and AI crossover space, where projects are building decentralized AI agent frameworks and tokenized compute networks, should take particular note. Many of these protocols rely on usage metrics and token throughput as indicators of network health and adoption. If Amazon’s own internal deployment shows how easily those numbers can be inflated when incentives are misaligned, decentralized AI networks with financial incentives baked directly into usage are even more susceptible to the same dynamic.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.