Harness Engineering: Why the Harness Matters as Much as the Model

A model is only half of an agent system. The harness supplies context, tools, permissions, state, feedback loops, and stopping conditions. When the same model performs differently under different runtimes, the missing variable is often the harness.nnA production harness should answer six questions: what can the model see, which tools can it call, what is the budget, how does it recover from failure, when must it stop, and how is the result tested? Strong harnesses turn a chat into a bounded engineering workflow.nnThe practical starting point is a tool map, a risk map, and an acceptance checklist. Future agent evaluations will increasingly measure task completion under constraints rather than one-shot answers.

Sources

See the linked primary sources in the Chinese edition.