A harness can hurt a frozen agent and help it after retraining
Your Policy Is Secretly a Harness: The Decision Geometry of Agent Interfaces · arXiv:2610.37255
In plain words
An agent harness is the prompts, tools, and rules that shape how a language model handles a task. This paper presents a method and theory for separating what a harness could make possible from how well a particular trained model uses it.
In a web-task test, changing how the observation was presented hurt a frozen model even though the best decisions stayed the same. In another test, a new harness first hurt the model, then helped it do better after further training. These results apply to the tested models and tasks.
The research
When a language-model agent performs worse after a change to its prompts, tools, or control logic, is the revised setup less capable, or does the trained model simply fit it poorly?
A frozen-model comparison cannot by itself distinguish those explanations, yet the answer changes how we interpret a harness revision.
The method separates a harness’s best possible return from a model’s mismatch, then tests how retraining changes the value of a revision.
The authors model how an agent harness shapes decisions and return, then separate a harness’s potential from a trained model’s fit. This gives a way to ask whether a revision that lowers a frozen model’s score might still pay off after further training.
In MiniWoB++, an H0-to-H1 change lowered frozen-model success, but the H1-adapted branch surpassed both the original H0 checkpoint and an equally updated H0 continuation on exact task completion.
This supports the account that adaptation can reduce mismatch enough to realize a harness’s higher potential.
The formal results assume finite states and actions, bounded rewards, and other shared comparison conditions.
The coding-agent estimates use a stronger model as a stand-in for an optimum, based on three probe tasks; the study does not establish that model is optimal or that the estimates generalize to other systems.
What the evidence shows
- What a harness comparison measures
The paper introduces a method and theory for analyzing agent harnesses, the prompts, tools, and control logic that present tasks to a language model and execute its outputs. It asks whether a harness comparison using one fixed model measures the harness’s potential or the model’s fit to it.
The distinction matters because a trained model may carry preferences learned under its original harness.
- Separate potential from fit
The theory separates deployed return into two parts: a harness’s ceiling, or the best return any policy can achieve under it, and mismatch, the return a particular policy loses through its decisions. Deployed return equals ceiling minus mismatch.
Training under a fixed harness changes mismatch, while a harness revision may change both parts, so a frozen-model score alone can mix them together.
- Same best decisions, different behavior
A presentation control tested a frozen Qwen3.5-2B checkpoint trained under H0. Reversing the observation fields while keeping their contents and the task mechanics fixed left the optimal decisions unchanged.
Yet the model changed its most probable action in 52% of states, and exact task completion fell from 85.0% to 19.2% on the fixed MiniWoB++ test set. For this tested change, the interface affected how the model expressed its learned preferences.
- A revision pays off after training
The H0-to-H1 test added ordered pairs of clicks that the agent could execute in one decision. With frozen weights, success fell from 85.05% under H0 to 72.90% under H1.
After 1,024 further updates under H1, success reached 88.17%, above both the source checkpoint under H0 and an equally updated continuation under H0, which reached 83.37%. The revision that hurt the frozen model helped after adaptation in this test.
- Testing predictions beyond one swap
Before adaptation, the authors’ ceiling-and-mismatch estimates ranked later gains across eight qualifying harness swaps more closely than frozen return changes did. A separate coding-agent study found that teacher-based estimates ordered five harnesses as their measured accuracy did, although the teacher was not known to be optimal.
Together, these results support evaluating a harness’s potential separately from a particular model’s fit, within the tested settings.
Deep
- Original paper authors
- Jian Mu, Ruibo Shen, Yao Shu
- Submitted to arXiv
- License
- CC BY 4.0