An agent harness is the operating layer that turns a language model into a working AI agent. The model supplies reasoning and generation; the harness decides what instructions and tools are available, passes results back into the loop, stores state, applies permissions, records traces, and decides when the run should stop or ask for help.
This distinction explains why the same model can behave very differently in two products. One harness may expose only a search tool and require approval for every action. Another may give the model a codebase, a terminal, persistent memory, automated tests, and room to retry. Model quality matters, but so do the environment and feedback loops wrapped around it.
The core pieces usually include a system prompt, tool calling, context management, an execution loop, error handling, and guardrails. Longer-running systems may add checkpoints, resumable state, specialist agents, evals, and detailed audit logs. Managed agent platforms, including the enterprise approach described in OpenAI's AWS expansion, package much of this plumbing for teams.
Why it matters at your desk. When an agent in Cursor succeeds at a refactor, the model is only part of the result; repository instructions, file tools, tests, and permission boundaries also shaped the work. The same is true for shared Workspace Agents in ChatGPT: connectors, approvals, schedules, and team controls determine whether the agent is useful in practice.
When comparing agent products, ask about the harness as seriously as the model. What can the agent access? How does it recover from failure? Where can a person intervene? Can you inspect what happened? Those answers often predict reliability better than a model name alone.