An agent harness is everything that is not the model: the loop, the tool definitions and dispatch, result formatting, context management, termination conditions, error recovery, escalation, and the evals that decide whether any of it is good enough. Discussion at the 2026 AI Engineer World’s Fair described teams shifting toward writing the systems that write the feature code, with engineers supervising harnesses rather than typing the integration by hand.
Worked example: most “the model is bad” reports turn out to be harness bugs — a vague tool-parameter description producing confidently wrong arguments, or an unbounded history quietly degrading quality as the conversation grows. Gotcha: frameworks hide exactly these decisions, which is why interview loops for agent roles frequently ban them.