A language-model agent is defined by both its model and its harness, the executable program that organizes model calls, tool use, and information flow. Different tasks require different harnesses, so adapting the harness at test time can improve performance without changing model parameters. The authors introduce harness learning, which trains a proposer model to revise the harness based on execution feedback. This is formulated as meta-learning over programs, where the proposer is trained via using the task performance of the revised harness as reward. At test time, the proposer iteratively refines the harness using only execution feedback, with no parameter updates. Experiments on reasoning and multi-hop question answering show that the trained proposer improves harness quality and that this ability transfers to unseen tasks. Policies trained on individual revisions can continue improving over multiple rounds, though the benefits of training on revision sequences vary. This work suggests a path toward agents that continually learn to improve their own workflows from experience.
Agents Learn to Revise Their Harnesses on Unseen Tasks
A reinforcement-trained proposer improved executable agent workflows from execution feedback, without updating model parameters at test time.
Independent
Alvin Zhang · Xuecheng Liu · Zixuan Wang · Fahim Tajwar · Daman Arora · Ruslan Salakhutdinov · +3 more
Research Digest··2 min read
Thread:Agent Harness Optimization
In one line
Harness learning trains a proposer to revise a solver's code-based harness using execution feedback, enabling test-time adaptation that generalizes to new tasks.
Before you read this
Three ideas this paper builds onagent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.186 papers →harness evolutionIteratively adapting an agent's environment, prompts, or evaluation setup to improve task performance and generalization.69 papers →test-time adaptationAdapting a deployed model during inference using new data or interactions, without a separate training phase.84 papers →
Read it as
Zhang et al. train a proposer model to edit the executable harness that controls a solver model’s calls, tools, and information flow.
The authors frame harness adaptation as meta-learning over executable programs.
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.