Harness adapters target model performance and cost
The approach pairs a trained adapter with an agent harness, according to a Hugging Face post.

Key takeaways
- A custom adapter can improve performance for a specific agent harness, the post says.
- The post reports lower token use after training.
- The training used reinforcement learning inside the harness.
- The source post is unverified.
In this article
What changed
Hugging Face has published a post about harness adapters. The post describes a method for adapting a model to an agent harness. The company says this can improve performance and reduce cost.
The post was published on 20 November 2026. The details come from the company and are not independently verified.
Important details

The post says the method uses reinforcement learning. Reinforcement learning is a training method where a model improves through rewards. The post says the training happens inside the chosen harness. The harness runs as it ships.
A proxy records the tokens that the model produces. The trainer then learns from those tokens. The company made ready-to-use images for the tasks. The post offers a reusable recipe for others.
The post reports tests on a model called Qwen3.8-27B. It says the model scored 39% with one harness. A custom adapter raised the score, the post claims. The post also says token use fell.
The company says the adapter made the model reason less and call fewer tools. The source post is unverified. The post mentions a super adapter that runs under several harnesses.
Who is affected
The approach may matter to developers who build agent tools. It may also matter to teams that use open models. The post says the recipe can be reused. The company says the work was done on the Hugging Face Hub.
The next steps are not clear. The post asks an open question about what the model learned. The company has shared the adapters and tasks.
What it means
The post suggests that custom training can help a model fit a real task. The company says the method can improve performance and cost. The source post is unverified. The company has not said how widely the approach will be used.
Sources
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



