autoharness vs harness2
Self-learning skill layer for Claude Code: distills skills from your real sessions, merges same-scenario ones, updates them in use and prunes the unused — touching only skills it wrote. — versus — Google Research's recursive harnessing framework: an agent refines its own harness (prompts, tools, memory, skills) per task via Propose–Probe–Compose, with model weights fixed.
Both evolve the harness from real runs. autoharness distills Claude Code skills; Harness² is a general research framework across models.
| autoharness | harness2 | |
|---|---|---|
| Stars | 11k | 23k |
| Forks | 652 | 2.3k |
| Language | Python | Python |
| License | MIT | Apache-2.0 |
| Last activity | 2 days ago | 4 days ago |
| Topics | skills, coding | agents, evals |
| Curated connections | 6 | 3 |
autoharness — the curator's take
Install it if you live in Claude Code and want lessons captured without curating skills by hand — zero config, zero deps, and it never touches skills you wrote or installed. Skip it off Claude Code, or if you need measured gains: skills survive on adherence, not a held-out score. Needs Python 3.11+ as `python3` on PATH, or its hooks stay off (macOS's stock /usr/bin/python3 is 3.9).
harness2 — the curator's take
Use it if you research or tune agent harnesses and want the agent to improve its scaffolding on unseen tasks without ground-truth labels. Lessons carry forward between tasks. It is a research framework: expect paper-grade code and extra model calls per task for probing. Skip it if you want a ready coding agent or a production orchestration runtime.