AutoHarness: open-source self-learning skill layer for Claude Code lifts CORE-Bench 42%→78%
thisguyknowsai · x · 2026-10-27
A new open-source tool, AutoHarness (MIT, v0.5.3), adds a self-learning skill layer to Claude Code: it watches your real coding sessions, distills lessons into skills, and automatically creates, merges, updates, and archives them — no manual SKILL.md writing.
Key capabilities:
- Writes new skills from work you already did; /learn saves a lesson on demand
- A reflector folds same-scenario skills into one instead of stacking near-duplicates
- Rewrites skills when you correct it and archives unused ones
- Opens each session with an index of what it learned
- Only touches skills it wrote itself, never user-installed ones
The README cites CORE-Bench (HAL): the same model with a different harness jumps from 42% to 78%, arguing the skill layer can maintain itself across model generations. Requires Python 3.11+; runs on Linux/macOS.
More from coding & agent
- Cheap verifier matches frontier LLM on RAG claim checks at 187x lower cost, but passes 23.2% of unsupported claims — rohanpaul_ai · 2026-10-27
- From 5-Word Prompt to Physical Product: AI Designs and 3D-Prints a Filter Holder — blixt · 2026-10-27
- "The agent understood me" isn't a product requirement: put implicit decisions in the spec — shoumikgoswami · 2026-10-27
- Claude Code + Codex re-tested 8 famous research findings overnight — only 1 held up — beglen · 2026-10-27
- One word 'fricative' unlocked full IPA planning: why domain knowledge still wins with coding agents — mattpocockuk · 2026-10-27
- From 126 review comments to zero: letting agents auto-merge PRs — calabi_and_yau · 2026-10-27