RiseHarness iteratively improves terminal-agent harnesses with weights fixed
Original title: Harness Self-Evolution for Long-Horizon Terminal Agents
It studies failed versus successful runs of the same task and adapts shared interaction mechanisms, not model weights.
RiseHarness iteratively revises the harness of long-horizon terminal agents while model weights stay fixed. It targets tasks that both succeed and fail under one configuration, using successful runs plus runtime-monitor traces to analyze command execution, feedback, context, and completion decisions. A Harness Handbook maps that evidence to shared mechanisms and source code, then pulls design references from other harness frameworks. Candidate changes are behavior-checked and re-evaluated; measured gains seed the next iteration.
- Stochastic errors can propagate through actions and context across repeated runs.
- The unit of adaptation is a specific model, not a universal harness.
- Ablations cover harness execution evidence and external research.
- Transfer is tested on unseen benchmarks with the same model.
From the original article
1 min read at huggingface.co
Long-horizon terminal agents may solve a task in one run yet fail across repeated runs. Stochastic errors can propagate through actions and context, creating a gap between capability and reliable execution. We present RiseHarness, a framework for iteratively improving agent harnesses with model weights fixed.
By Ruhan Wang, Yucheng Shi, Zongxia Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.