Ecdysis paper speeds self-evolving agent harness training 1.84x, lifts reasoning accuracy 18.56%
dair_ai · x · 2026-09-12
dair-ai highlights the Ecdysis paper, which tackles two practical problems of self-evolving agent harnesses: slow search (each candidate harness needs repeated agent runs and code edits) and overfitting fixes (each failure is patched as if it were a harness bug even when the model caused it).
Ecdysis's approach:
- Analyzes failures across a batch of tasks and repairs only recurring patterns
- Multiple diagnostic roles agree on a change specification before any code is modified
Results:
- Harness training runs up to 1.84x faster than existing harness evolution methods
- Resulting harnesses gain 18.56% in reasoning accuracy
- Better transfer across LLMs and fewer tokens
- Matches full-data training with a quarter of the data
More from coding & agent
- GitHub Copilot App uninstall leaves MCPs, plugins and old versions behind — PaulShellDev · 2026-09-21
- Dev uses small model to triage resumes, cutting tokens so much it's given away free — draginol · 2026-09-21
- HF researcher: doing is harder than criticising — agent workflows should leverage both roles — antoine_chaffin · 2026-09-21
- The coding AI consensus has flipped: the crowd that moved to Claude Code is now on Codex — BLUECOW009 · 2026-09-21
- TypeSafe's Jev Cut Grok Build Costs 22-40% on the Same Tasks — Daniel_Farinax · 2026-09-21
- jev-skill-suggester: a lightweight pre-execution skill router stops agents picking wrong tools — udmrzn · 2026-09-21