RiseHarness improves terminal-agent harnesses with model weights fixed

What the research found

RiseHarness is an arXiv paper (2610.27435) that iteratively improves terminal-agent harnesses with model weights fixed, using successful and failed runs as references and adapting mechanisms from other frameworks.

Claims and how well the evidence supports them

1. RiseHarness improves agent harnesses without altering model weights 99%
2. RiseHarness uses successful and failed runs under the same configuration as references 91%

Evidence

• huggingface.co https://huggingface.co/papers/2610.27435

Actionability rated low.