Harness Learning: RL-Trained Proposer Adapts Agent Harnesses at Test Time Without Weight Updates

pmddomingos · x · 2026-10-12

Pedro Domingos shared an arXiv paper, Harness Learning Enables Generalizable Test-Time Adaptation, quipping "the model swallows the harness."
The paper frames an agent as model + harness (the executable program organizing model calls, tool use, and information flow), and trains a proposer model to revise a solver's harness from execution feedback.
It's formalized as meta-learning over executable programs, with harness revisions playing the role of weight updates; the proposer is trained with RL using revised-harness task performance as reward.
At test time, the proposer refines the harness from successive executions on new tasks with zero parameter updates. Experiments on reasoning and multi-hop QA show better revisions and transfer to unseen tasks, pointing toward continually learning agents.

More from coding & agent