AHE paper auto-evolves coding-agent harnesses with decoupled files, lifting Terminal-Bench 2 to 77%
morgymcg · x · 2026-10-20
Reading Google Cloud's Harness2 paper, the author highlights that its agent builds on AHE (Agentic Harness Engineering, Fudan/Peking Univ.), which uses "loosely decoupled" components — system prompt, tools, middleware, skills, sub-agent config, long-term memory — exposed as explicit files at fixed mount points, so edits stay localized with git-level diffs and rollback.
Key results: 10 AHE iterations lift Terminal-Bench 2 pass@1 from 69.7% to 77.0%, beating the human-designed Codex harness (71.9%); the frozen harness transfers to SWE-bench-verified with 12% fewer tokens. Ablations attribute gains to tools, middleware, and memory rather than the system prompt. The author likes the simplicity of the design.
More from coding & agent
- Decisions API integration with Live API showcased as fast GPT backend in demo — gabrielchua · 2026-10-20
- OpenAI ships Decisions API integration with the Live realtime voice API — gabrielchua · 2026-10-20
- DoorDash runs 130K automated tasks a month: inside 8 AI software factories — Pavan_Belagatti · 2026-10-20
- Agents Ask for Approval 20 Times a Day—Are You Still Reviewing or Just Clicking Yes? — Glittering812 · 2026-10-20
- Skyrim AI mod lets bards compose songs about your adventures — and got flamed off Reddit — ReignMocap · 2026-10-20
- LangChain's Harrison Chase: how vertical agents differentiate via context, harness and model — hwchase17 · 2026-10-20