• Build agent architectures that are fast, reliable, and actually pleasant to work with • Research and ship techniques for agent reasoning, planning, and tool use • Create evaluation pipelines so we can measure whether agents are getting better or worse • Talk to users. Understand what's broken. Fix it.