Harness Engineering: How AI Agents Actually Work

Most AI agent systems make the same mistake: they give the model too many tools at once. I break down how harnesses actually work, from why bash-style interfaces are so good for tool discovery to how context bloat hurts agents. I also explain why the KV cache matters for speed and cost, and walk through the harnesses I've built, including coding agents, browser automation, OpenChamber, and the early ideas behind Pinkzap. The chapters also cover compaction, dynamic tools, latency, and why coding harnesses are mostly solved.