Self-Harness: Shanghai AI Lab's Method Lets Agents Fix Themselves, 30.2%→84.6% Without Retraining
Promptmethus · x · 2026-11-05
Shanghai AI Laboratory published a free paper, Self-Harness, describing a method for AI agents to repair their own setup — no engineer or stronger model needed. The agent reads its own failed runs, identifies recurring breakage patterns, and writes its own rules. Same model, same weights, zero retraining: success on unseen tasks jumped from 30.2% to 84.6%, the kind of jump consultants charge $3,000–10,000 for.
The method boils down to four steps:
- Pull your last 20 sessions, failed ones first
- Identify the 3 failures that keep repeating
- Write one small rule per failure — never a big rewrite
- Rerun old tasks; a rule stays only if it passes
This also explains why copying someone else's CLAUDE.md never works: your agent has its own bad habits, so rules must come from your own failures. The thread includes a ready-to-paste Claude Code prompt template.
More from coding & agent
- 5 ways to connect AI agents across frameworks with the A2A protocol (26.4k stars) — goyalshaliniuk · 2026-11-05
- Ask your model what's wrong with your harness — an easy trick to catch silent bugs — insomniating · 2026-11-05
- "Haven't pushed in 6 months, it's all AI code and worth nothing" — the wildest take in the Claude Code repo-wipe thread — ryunuck · 2026-11-05
- Solo dev builds hyper-realistic browser jet simulator with Claude Opus 5.5, 25,000 players in two weeks — aCakeShip · 2026-11-05
- FrontierPhysics: 56 expert tasks benchmark AI agents on end-to-end physics research — Bingran You · 2026-11-05
- Free browser tool turns iPhone Duo screenshots into 3D fold animations up to 4K — yihui_indie · 2026-11-05