What makes a harness a harness: necessary and sufficient conditions for an agent harness

The paper presents an operational definition of "agent harness," clarifying its necessary and sufficient conditions as a layer that enables a language model to function as a coding agent within software repositories. It distinguishes the term from related concepts like agent frameworks and evaluation harnesses, and applies this definition to analyze six existing systems, including Claude Code and Codex CLI. This work provides a structured vocabulary and framework for practitioners, facilitating clearer communication and comparison in the development of agentic systems in AI.