Harness-S: rescuing failed agent rollouts via sibling harnesses into one policy
_akhaliq · x · 2026-11-06
The paper Harness-S: Learning from Sibling Harnesses trains agents by exploiting complementary "sibling" harnesses:
- Failed agent rollouts often stem from harness (orchestration) flaws, not the policy itself; different harness variants have complementary strengths.
- A pool of harnesses is evolved offline with root-cause-directed repairs; when all rollouts under base harness H₀ fail, an adaptive router picks Top-K sibling harnesses whose successes replace the base rollouts in the GRPO update.
- The rescued behaviors transfer into a single policy that still works under the original harness H₀, and pairing with a Rank-1 harness yields further gains.
Key idea: turn "works if you swap the harness" into a durable capability of the policy itself.
Related event: Harness-S: Rescuing Failed Rollouts with Sibling Harnesses(2 posts)→
More from coding & agent
- 'Software factory' is just marketing: good engineering is a kitchen where agents cook and engineers run the pass — 0xkarasy · 2026-11-07
- The AI discount trap: why demanding cheaper AI-assisted development can cost you more — luisdans · 2026-11-07
- Founder wires an agent to check his product's ranking every 10 minutes and ping Discord — pramodk73 · 2026-11-07
- Two engineers argue whether prompting still matters in the agent era, with a pipeline-fix prompt example — teodorio · 2026-11-07
- Hermes-Agent hacks Echo Show 5: screen is just a web page you can customize — alexcovo_eth · 2026-11-07
- 5 vibe coding component libraries worth pasting into Claude or Cursor — thisiskp_ · 2026-11-07