
The discussion centers on long-horizon agents, agent harnesses, and the evolving landscape of AI development. It features Harrison, presumably an expert in the AI field. The conversation highlights the recent advancements in making long-horizon agents functional, particularly in coding and AISREs, where they can produce first drafts of work. A key point is the distinction between models, frameworks, and harnesses in agent construction, with harnesses offering more "batteries included" features like planning tools and compaction methods. The speakers explore whether coding agents represent a subcategory or a general-purpose AI and also examine how building agents differs from traditional software development, emphasizing the importance of traces and iterative processes.
Sign in to continue reading, translating and more.
Continue