How It Works
Agent Harness
Everything wrapped around an AI model to turn it into something that can actually do a job: the loop that lets it take one step after another, the tools it is allowed to call, the memory of what it has already done, the sandbox it runs inside, and the limits and approvals that stop it running away with your money. The usual shorthand is “Agent = Model + Harness.”
Origin · no single documented coiner
No single person coined “harness” in this sense — it was inherited from the older software “test harness” and from machine-learning “evaluation harnesses” such as the one used to score SWE-bench, then stretched to cover the whole wrapper around a model. The UK’s AI Security Institute was already describing an agent as a model plus its scaffolding in 2023. The name for the discipline, “harness engineering,” is credited to Viv Trivedy, whose “Anatomy of an Agent Harness” post Addy Osmani points to as the clearest derivation. Usage is still loose: people call Claude Code, Codex CLI and an evaluation script all “harnesses,” which is why a 2026 arXiv paper set out to define the boundary.