paper
1 articles
Natural-Language Agent Harnesses: When an Agent's Soul Moves from Code to Plain Text
A Tsinghua Shenzhen team proposes NLAH (Natural-Language Agent Harnesses): moving agent control logic from code into structured natural language, executed by an IHR runtime. Experiments show harnesses can reshape agent behavior patterns entirely, but more structure doesn't always mean better results. Dan McAteer argues harness engineering matters as much as model capability.