HarnessPro lets AI agents rewrite their own harness to cut task cost
Overview
Researchers from Tsinghua, Zhipu, Shanghai Jiao Tong, University of Warwick and Kuaishou release HarnessPro, a method that lets an AI agent improve the harness it runs in.
The paper was posted to arXiv on October 19, and the code is on GitHub. The authors add a second goal to self-improvement: raise performance while lowering execution cost.
Pandaily reports that on Terminal-Bench 2 with GPT-5.5 and Codex, the Heavy variant raises pass rate from 76.3% to 80.8% and cuts cost per task from $1.09 to $0.72, compared with the RHO method. The cost figures cover running tasks only, not the search and evaluation used to produce the harness.
Written by AI from the articles below · updated Oct 31, 3:53 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- PandailyNewsTsinghua and Zhipu's HarnessPro lets agents rewrite their own harness to cut task costs
AIResearchers from Tsinghua, Zhipu and partner institutions released HarnessPro, a method that lets an AI agent improve its own harness for performance and lower cost. On Terminal-Bench 2 with GPT-5.5 and Codex, the Heavy variant raised pass rate from 76.3 percent to 80.8 percent while cutting cost per task from $1.09 to $0.72 compared with the RHO method. The cost figures cover running tasks, not the search and evaluation behind them.
Heat Trend
Not enough continuous observations to show a trend yet.