Agent harness

The software that runs an AI model as an agent. It supplies tools and context, enforces permissions, records state, and decides what happens after a tool error or model response.

An agent harness is the runtime around an AI model. It gives the model access to selected tools and context, validates tool calls, enforces permissions, records what happened, and decides whether a failed step should retry, stop, or take another route. The model proposes actions inside that environment; the harness determines which actions are available and how the application carries them out. Two products can use the same model and behave very differently because their harnesses expose different tools, instructions, state, and recovery rules.

Builder example

Model capability is only one source of production behavior. A temporary API failure becomes a short retry or a broken workflow depending on the recovery rules. A tool can reach a private record only if the permission layer allows it. A bad result can be diagnosed only if the run leaves enough state and logs to reconstruct what happened. Model choice and harness design therefore need separate evaluation: one test asks whether the model can do the task, while another asks whether the surrounding system runs that capability within the intended limits.

Common confusion: A harness does not guarantee that an agent is safe or reliable. Its rules can be incomplete, its validators can accept the wrong arguments, and retries can repeat a harmful action. The term names the control layer around the model, not a certificate that the controls are correct.