Hugging Face has published a method for training a small add-on to an open-weight model that is tuned to a specific agent harness, the scaffolding that wraps a model so it can call tools and complete tasks. The approach, which it calls harness adapters, uses reinforcement learning inside the harness itself, on a custom task distribution, and produces a LoRA adapter while the base model weights stay frozen. In the example given, Qwen3.8-27B scored 39% pass@1 on Terminal-Bench 2.1 with the pi harness, using an average of 2.9K tokens per task. With a custom RL-trained LoRA adapter added and the weights otherwise frozen, the same model reached 55% pass@1 while averaging 1.5K tokens. Hugging Face attributes most of the gain to the model reasoning less, calling fewer tools and getting tasks right on the first attempt. Earlier experiments on data-analysis and code-building tasks had shown similar benefits. This work tests whether the result holds at 27 billion parameters on harder real-world software engineering tasks, with one adapter trained for each of three harnesses: pi, OpenCode and Hermes Agent. The post presents the work as a reusable recipe so developers can reproduce it for their own task, harness and model combinations. The article as provided was cut off before the full methodology and complete results, so the scope of the evaluation beyond the headline figures cannot be confirmed from the text. For Australian businesses and professionals, this is research-stage work rather than a product change. Training an adapter requires GPU compute and RL tooling, and the token reduction only translates into savings for teams that run open-weight models themselves or pay per token for hosted inference. Teams that place customer or staff data, including personal information in agent transcripts, into training or evaluation should check their obligations under the Privacy Act 1988. For most, nothing changes immediately. Teams already running open-weight coding agents could reasonably benchmark the approach on their own tasks.