What changed

Hugging Face published a reproducible recipe and adapters that reinforcement-train a Qwen3.8 27B base model inside pi, Hermes and OpenCode harnesses. It reports Terminal-Bench 2.1 rising from 39% to 55% while average output falls from about 2,900 to 1,500 tokens in the target harness.

Why it matters

The result suggests harness behavior is part of the model system and can be optimized directly, offering a cheaper alternative to treating prompts, tools and model weights as independent components.

What to do

Reproduce one adapter with the published datasets and jobs, then evaluate it in both target and non-target harnesses for success, token use, transfer loss, tool errors and safety regressions.

Caveat

The reported gains are Hugging Face results on a specific base model and benchmark. Training used two B200 GPUs, and harness specialization can reduce portability or overfit evaluation behavior.