The Harness Tax: The Dead Weight Inside Your Coding Agent

A benchmark comparing the token overhead of three AI coding agents (Claude Code, OpenAI Codex, and Pi/OpenClaw) reveals a hidden cost called the 'harness tax' — the tokens an agent spends on itself before doing any useful work. Claude Code consumed ~27,000 input tokens per request vs Pi's ~2,600 for the same trivial task. Over a 40-turn session, this compounds to over 1 million wasted input tokens. Beyond cost, harness bloat causes 'context rot,' degrading model attention on actual task content. The post argues that as models improve, thick harnesses become dead weight, and the better design pushes complexity into the model and infrastructure rather than the context window.