An AI agent harness provides the operational structure around a model. It can manage system instructions, tool schemas, permissions, memory, workspace access, approval requests, retries, output handling, and integration with terminals, editors, browsers, or external services.
A harness influences what an agent can accomplish and how safely it operates. Two agents using the same model can behave differently when their tools, context, policies, or lifecycle controls differ, so evaluations should identify the harness and configuration as well as the model.
ELI5
An agent harness is the software wrapped around an AI model that lets it do useful work. It supplies instructions, tools, project context, permissions, memory, and rules for handling failures or asking a person before sensitive actions.
For example, the same model might only answer questions in one harness but be able to edit files and run tests in another. The harness strongly affects what the agent can do, what information it sees, and which safety boundaries it must follow.
