The authors factorize the agent harness space into five components: system prompt, tool interface, observation, context management, and workflow.
Harness diversity boosts agentic RL generalization across execution configurations
Training reinforcement learning policies across varied harness configurations improves transfer to unseen white-box and black-box harnesses.
Top University
Fei Bai · Huatong Song · Chuan Hao · Ran Tao · Jian Yang · Bryan Dai · +2 more
IQuest Research · Beihang University · Renmin University of China
Research Digest··3 min read
The authors introduce Harness Scaling, a method that trains agentic reinforcement learning policies across multiple execution harness configurations.
§