Designing the full environment and feedback system around an agent.
A harness combines context assembly, tools, permissions, execution, state, observability, evaluation, and feedback loops. Its purpose is to make capable model behavior repeatable and bounded under real operating conditions.