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

Researchers provide a formal operational definition of 'agent harness' in AI software engineering, establishing necessary and sufficient conditions to distinguish harnesses from related tools like frameworks and SDKs. The work analyzes six real-world implementations and proposes a shared vocabulary to standardize how the industry discusses and compares agentic systems built on language models.

This academic work addresses a fundamental problem in emerging AI infrastructure: the lack of precise terminology around agent harnesses, which have become central to how language models interact with code repositories and perform autonomous tasks. As tools like Claude Code and OpenHands proliferate, developers and researchers use 'agent harness' inconsistently, sometimes referring to complete products, evaluation scaffolds, or underlying frameworks interchangeably. This semantic ambiguity creates friction in engineering communication and makes scientific comparison of agentic systems difficult.

The research traces the genealogy of 'harness' terminology from equestrian tack through software testing to contemporary AI applications, grounding the definition in historical precedent. By combining conceptual analysis with documentation from primary sources, the authors establish clear boundaries separating agent harnesses from adjacent concepts like IDE plugins, orchestrators, and evaluation frameworks. They validate their definition against six prominent implementations, demonstrating practical applicability.

For the AI development community, this work provides immediate value by offering a shared vocabulary that facilitates clearer technical discussions and enables apples-to-apples comparison of competing solutions. Better semantic precision supports more rigorous scientific research on agentic systems and helps standardize how practitioners evaluate tool selection. The operational definition also guides engineers designing new harnesses by clarifying what architectural components are essential versus optional.

Looking forward, this contribution establishes a foundation for deeper investigation into design tensions within agent harnesses—questions about extensibility, transparency, and integration that will likely shape the next generation of AI development tools. As agentic systems become more prevalent, standardized terminology becomes increasingly valuable for both academic research and commercial development.

• →The paper establishes a formal, testable definition of agent harness by identifying necessary and sufficient architectural conditions.
• →Semantic clarity enables better scientific comparison and engineering practice across the growing ecosystem of agentic AI tools.
• →The definition successfully distinguishes agent harnesses from frameworks, SDKs, plugins, and orchestrators through consistent inclusion/exclusion criteria.
• →Real-world validation against six implementations (Claude Code, Codex CLI, Aider, Cline, OpenHands, SWE-agent) demonstrates practical applicability.
• →Standardized vocabulary addresses a critical gap in AI infrastructure documentation that has grown as autonomous agents become mainstream.