The interesting part of an Agent Harness is what you add on top (www.martinrichards.me)

Recent discussions in the AI/ML community have refined the concept of an "Agent Harness," which encapsulates the skills, workflows, and methodologies used to instruct an AI agent. Viv Trivedy introduced a succinct formula: Agent = Model + Harness, clarifying that while models like Claude Code and Codex provide reasoning, the harness includes all supporting tools and configurations that enhance that reasoning. The significant revelation is the distinction between the mechanisms (like runtimes and execution hooks) and the content (skills and workflows) within the harness, allowing users to better understand how to configure AI agents effectively. As the definitions evolved, the idea of a harness has been split into two primary components: "Skills," which provide the necessary context and information to the agent, and "Loop," which dictates the process workflow the agent follows. This separation helps clarify how users can tailor AI agents to their specific needs by either enhancing the knowledge base (Skills) or refining the execution process (Loop). By embracing this refined terminology, practitioners can more effectively construct agents that meet their objectives, ensuring that both the structural and operational aspects are comprehensively addressed in their design.