"Harness Learning" Lets AI Agents Rewrite Their Own Scaffolding
The paper introduces harness learning, where a trained proposer model revises an LLM agent's harness—the program orchestrating model calls, tool use, and information flow—using execution feedback, framing harness revisions as analogous to gradient updates in meta-learning. The proposer is trained with reinforcement learning using downstream task performance as reward, and at test time it adapts the harness to new tasks purely through iterative execution feedback, with no parameter updates. Experiments on reasoning and multi-hop QA show improved revision quality and generalization to unseen tasks, with gains compounding over multiple refinement rounds in some settings.
Discussion: 1 tweets from 1 authors · @pmddomingos