JHU & CMU Paper: Small Models Can Learn to Rewrite Agent Harnesses That Generalize to New Tasks
rohanpaul_ai · x · 2026-10-06
Researchers from Johns Hopkins and Carnegie Mellon introduce "harness learning," where a small proposer model is trained via RL to rewrite an agent's harness code using execution feedback.
Key points:
- The harness — the executable program organizing model calls, tool use, and information flow — jointly defines an agent with its model, and different tasks need different harnesses.
- The process is formulated as meta-learning over executable programs, with harness revisions playing the role of weight updates in gradient-based adaptation.
- On reasoning and multi-hop QA benchmarks, harness learning improves revision quality and transfers to unseen tasks at test time, with no parameter-space updates.
- The results suggest a path toward continually learning agents that turn accumulated experience into generalizable improvements.
Related event: Harness Learning: 4B Model Rewrites Agent Code, Beats 35B Teacher(3 posts)→
More from coding & agent
- Should an AI call agent keep running when one tool goes down? A degraded-mode proposal — SubtleInterval_4449 · 2026-10-24
- Claude Desktop blocks local models; users bypass it with a loopback proxy trick — pneuny · 2026-10-24
- OpenAI ships Automations in Agents API for scheduled agent tasks — OpenAIDevs · 2026-10-24
- Arca runs its financial services business on long-horizon agents, not with them — AI Engineer · 2026-10-24
- AGENTS.md explained: one root Markdown file to rule every AI coding tool — adnan_hashmi · 2026-10-24
- Summarizer hallucinates details from headline-only feeds, dev asks how to check faithfulness — rahul_watertech · 2026-10-24