AI glossary · Basic · Agents · Language models
Harness Adapters
A harness adapter is a connecting piece that bridges an AI model and the harnesses it runs in (frameworks that give a model tools, prompts and an execution environment so it can do work), so the model works well across several of them. Because the term is still new, it sometimes means connection code and sometimes means a LoRA adapter trained separately for each harness.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Perception & action
- Fields
- Agents, Language models
- Added
- 2026-11-07
What it is
Building an AI agent means choosing not just a language model but also the harness in which that model acts. Each harness differs in how tools are called, how errors are reported and how prompts are structured. A harness adapter bridges these differences so the same model can be plugged into multiple harnesses, or made to fit a particular harness better. The meaning of the term has not settled yet, so it is best to check which sense each article uses.
How it works
Its uses appear to fall into two broad categories. The first is software connection code. The paper "Finding the Right Fit," posted on arXiv in October 2026, released harness adapters built to test multiple models in multiple harnesses, along with its evaluation code. On October 29, 2026, there was also a report that Infini-AI-Lab released Hermit, an open-source harness adapter framework for AI agents. The second is trained adapters. One community post put forward, in its title, the idea of attaching a separate LoRA adapter (a method that trains only a small set of additional parameters rather than the whole model) to the model for each task and each harness.
Two meanings of harness adapter
Connection codeSoftware adapter
- Lets a model plug into multiple harnesses
- Does not change the model itself
- Used for fair comparative evaluation
LoRA adapterTrained additional parameters
- Trained separately per harness and task
- Makes the model fit a harness better
- Increases adapter maintenance burden
공통 Both are ways to improve the fit between a model and a harness
Why it is getting attention
It has become clear that the same model can score very differently depending on which harness it is placed in. The paper above tested 66 configurations combining 4 harnesses, 5 models and other factors, and reported that on Terminal-Bench 4, Claude led GPT by 7.94 points in the OpenHands harness but trailed it by 30.16 points in the PI harness. The harness from the company that built the model was not always the best, and higher cost did not necessarily mean higher scores. The researchers interpreted this as showing that, when a model tries to fix its own failures, what matters is whether the harness returns errors in a form the model can use, and argued that models, harnesses and tasks should be evaluated together.
Fit differences shown in the paper
Things to know (limitations)
The term does not yet have a settled definition in the industry, so you have to tell from context whether it refers to connection code or a LoRA adapter. There is also little source material to rely on: one paper plus some reports and community posts. With the LoRA approach, an adapter has to be trained and maintained for each harness, which increases the operational burden. Also, a combination that worked well on one benchmark may behave differently on other tasks, so it is safer to check directly with tasks similar to your own work.
Example
Suppose a team planning to adopt a coding agent in-house is weighing three candidate models. If each model is compared only with its official harness, harness differences can look like model differences. So the team uses a harness adapter to plug each of the three models into the same two or three harnesses, tests them on internal tasks, and picks a 'model + harness' combination by looking at both scores and cost per task. The adapters and 6,204 scored trajectories released with the "Finding the Right Fit" paper serve as reference material for this kind of comparison.