AI coding agents face significant noise pollution from development tools flooding context windows with irrelevant output. A TypeScript monorepo using Turbo dumps approximately 1005 words of build data per package, creating substantial token waste. This distraction reduces coding agent effectiveness, similar to how dogs perform poorly when surrounded by environmental stimuli during training.

The author proposes implementing an LLM=true environment variable that would prompt tools to automatically reduce output like colors and progress indicators. This approach would create a win-win-win scenario: reducing costs by burning fewer tokens, improving context window quality, and decreasing energy consumption. Current solutions involve manual configuration of various tool-specific flags.

As AI coding becomes the default paradigm, the article provocatively suggests we might eventually need to set HUMAN=true instead. The author calls on developers like Boris Cherny to implement the LLM=true variable in Claude Code by default. This small change could collectively save substantial resources across the development ecosystem without requiring changes to existing tool behavior.