Code Harness vs Natural-Language Harnesses for AI Agents

This article compares two approaches to controlling AI agent execution: code-based harnesses versus natural-language agent harnesses (NLAHs), where orchestration logic including tool invocation timing, state management, validation, and recovery strategies is extracted from code and represented as editable natural language documents executed by an intelligent runtime. Evaluation across three benchmarks — SWE-bench Verified, Terminal-Bench 2.0, and OSWorld — shows that natural language harnesses achieve comparable performance to traditional code implementations while significantly reducing static policy complexity.