Natural-Language Agent Harnesses: Externalizing Agent Control Logic as Portable, Editable Artifacts

Summary

A new arXiv paper (2603.25723) by Linyue Pan, Lexiao Zou, Shuo Guo, Jingchen Ni, and Hai-Tao Zheng introduces Natural-Language Agent Harnesses (NLAHs) and the Intelligent Harness Runtime (IHR). The work addresses a key problem in NLP and agent research: agent performance increasingly depends on harness engineering, yet harness designs are typically buried in controller code and runtime-specific conventions, making them hard to transfer, compare, and study scientifically. The authors propose externalizing the high-level control logic of agent harnesses as portable executable artifacts. NLAHs express harness behavior in editable natural language, while IHR is a shared runtime that executes these harnesses through explicit contracts, durable artifacts, and lightweight adapters. The paper reports controlled evaluations across coding and computer-use benchmarks, covering operational viability, module ablation, and code-to-text harness migration. Published March 26, 2026, this research offers a new direction for making agent control frameworks portable and scientifically comparable.