In plain words
An agent harness is the surrounding software that turns model calls into a working agent system. It coordinates the model with tools, state, instructions, and an environment in which actions can happen.
A closer look
A harness may run the agent loop, deliver tool results, manage permissions, preserve progress, handle errors, and decide when to stop. It can also manage a limited context window by summarizing or retrieving information. Which responsibilities belong in the harness varies across systems.
The same model can behave differently in different harnesses because its available information, tools, and feedback differ. Improving the harness can improve task performance without changing the model’s weights. Separately, an evaluation harness is software that runs standardized tests and collects results; the shared idea is controlled execution around a component.
In practice
For a coding agent, the harness gives access to selected files, executes approved commands, captures test results, and feeds those results back to the model. It may save progress so the task can resume later.
A useful distinction
A harness is not another name for the model, and a prompt alone is not the whole harness. It is the operational software around the model, including behavior the model cannot enforce by itself.