Natural-Language Agent Harnesses (NLAHs)

Source: https://arxiv.org/abs/2603.25723 Author: Linyue Pan, Lexiao Zou, Shuo Guo, Jingchen Ni, Hai-Tao Zheng Date: 2026-03-26

Summary

Framework allowing developers to express agent control logic in editable natural language rather than hardcoded controller code. Introduces Intelligent Harness Runtime (IHR) to execute these harnesses through explicit contracts, durable artifacts, and lightweight adapters.

Key Claims

  • Agent harnesses are typically buried in controller code — hard to study, compare, transfer
  • NLAHs externalize control logic as portable natural-language artifacts
  • IHR executes harnesses through contracts — consistent execution
  • Improved transferability enables comparing harnesses as scientific objects
  • Evaluated on coding and computer-use benchmarks with module ablation studies
  • Code-to-text migration tested: existing harnesses can be converted to natural language

Connection to Other Sources

Directly related to Anthropic’s effective harnesses — both about the harness abstraction layer for long-running agents. This paper makes harnesses portable and editable.

Concepts