Coding agent harness
The tools, workspace, limits, checks, and feedback loops around a coding agent's model that turn raw skill into steady work.
A model on its own does nothing useful in a repo. The harness is everything wrapped around it. The place the code runs. The tools it may call. The files it is shown. The checks that run on its output. The loop that feeds results back in. Two products can use the same model and still act nothing alike, because their harnesses differ.
Through 2025 and 2026 the field's focus moved. First prompt wording. Then context contents. Then the harness itself. Some call this harness engineering. The reason is simple. An agent's reliability comes mostly from the limits and feedback loops around the model. The phrasing of the request matters far less.
A good harness makes failure cheap and easy to see. Work runs where it cannot break anything. Machines check the output instead of vibes. The agent sees its own errors before a human ever does.