Building Deep Research: How we Achieved State of the Art

AI research agents are an important application for collecting, reading, and synthesizing information at scale. Building a state-of-the-art agent requires a flexible software layer, or "agent harness," that can adapt to future model improvements by simplifying orchestration and leveraging emerging capabilities. A key challenge is "context engineering," which involves carefully curating the agent's context window to maintain focus during long tasks. This approach models human research by having the agent distill information into reflections to guide its next steps, which is more efficient than traditional methods that accumulate raw data in the context window.