"A Great Model with a Bad Harness Loses": Complete Guide to Harness Engineering
aakashgupta · x · 2026-10-17
Aakash Gupta published a long-form guide on harness engineering, anchored in Addy Osmani's line (857K views) — a former Chrome lead now working on Claude Code: "A decent model with a great harness consistently beats a great model with a bad harness."
The article frames the harness as the half of an agent you actually control (Agent = Model + Harness), a superset of prompt and context engineering. The cautionary tale: in August, Sebastian Guillemot's dev machine got wiped when Fable ran rm -rf on his home directory while testing a sandbox — exactly what a proper harness should have stopped.
Covered: what a harness is, 4 harness types PMs should care about, lessons from a Personal OS, product-building case studies, ownership and the Harness Spec, and why harness engineering keeps growing in importance. Gupta interviewed practitioners like Ryan Lopopolo and iterated his own harnesses while writing it.
Related event: Good Harness Beats Strong Model: Addy Osmani's Agent Engineering Thesis(2 posts)→
More from coding & agent
- Text-to-3D iOS games: Abacus AI Agent generates native 3D apps from one prompt — bindureddy · 2026-10-17
- Inside Limestone's AI software factory: from ticket to tested PR with agents — alex_verem · 2026-10-17
- Hermes agent spins up its own GitHub repo: local-first elder-care automation kit for Home Assistant — NousResearch · 2026-10-17
- Developers say Codex has become unusable and are switching to Claude — zealcaiden · 2026-10-17
- Qwen Code ships v0.25.3-preview.0 with a batch of agent stability fixes — qwen-code-review-bot · 2026-10-17
- "AI hasn't solved coding": easy tasks hide the bumps, warns dev in Anthropic debate — gerardsans · 2026-10-17