Improving Deep Agents with harness engineering

Harness engineering involves building systems and tooling around a model to improve its performance on specific tasks. The performance of a coding agent was significantly improved on a benchmark by only changing its harness, not the underlying model. This was achieved by using trace analysis to understand failure modes and then implementing strategies like self-verification loops to encourage testing and refinement. Providing agents with environmental context, such as directory structures and time limits, and strategically managing their reasoning compute budget also proved effective for improving autonomous performance.