Harness
also: harness engineering
The orchestration code around a model — the loop, tools, retries, budgets — that turns it into a product.
The harness is everything that runs the model: assembling context each turn, parsing outputs, executing tool calls safely, handling retries and errors, enforcing budgets and stop conditions, logging traces. Harness engineering is the emerging discipline of building this layer well — two systems with the same model can behave worlds apart because the harness, not the model, decides what actually happens.
Worked example: the engineering scaffolding around a model that turns raw next-token prediction into a reliable system — the tool-call loop, context assembly, retries, verification, and stopping conditions. Gotcha: the same model scores very differently by harness (SWE-bench results swing widely on scaffolding alone), so ‘the model can’t do X’ is often really ‘the harness didn’t give it the tools/feedback to do X’; much of applied AI progress is harness work, not model work.