Harness Engineering: How OpenAI and LangChain Got Agents to Ship a Million Lines of Code
blaizedsouza · x · 2026-10-23
Gaurav Sen's 12-minute read breaks down harness engineering: OpenAI had 3 engineers ship 1M lines of agent-written code, and LangChain jumped its coding agent from Top 30 to Top 5 on Terminal Bench 2.0 without changing the model — both by changing the code around the model.
Key points:
- Agent = Model + Harness: system prompts, tool descriptions, sandboxes, subagent handoff logic, and per-step hooks are all the harness
- Agents are unreliable by default: hallucinated paths, violated module boundaries, dead-code accumulation
- Three failure modes addressed: context-window memory loss on long tasks (memory bases), token-heavy tool calls (context engineering), and code bloat (architectural constraints + entropy management)
- OpenAI's framing: design the environment, specify intent, build feedback loops — "Humans steer. Agents execute."
More from coding & agent
- DAB, a cross-database agent benchmark, hits EMNLP Oral; best model only 38% pass@1 — adityagp · 2026-10-24
- 27B open-source model beats GPT-5.5 at computer tasks via evolved interfaces, at 1.5% the cost — alex_verem · 2026-10-23
- 27B open-source model beats GPT-5.5 and Opus 4.8 on desktop tasks at 1.5% of the cost — alex_verem · 2026-10-23
- Interactive 3D AI server explainer built in one prompt with Claude and the open-source colorfigures skill — Promptmethus · 2026-10-23
- poem turns Codex sessions into chats — approve agent tasks from your iPhone — patrickreza · 2026-10-23
- Watching AI test your app like a real, picky customer — randal_olson · 2026-10-23