What does it take to build an AI scientist in chemistry that can carry out research and improve through experience? In this talk, I will explore scientific autonomy as a process of hill climbing: identifying the capabilities that limit autonomous research, addressing those gaps, and evaluating progress on increasingly demanding chemistry tasks.
\r\nDrawing on the El Agente family of systems, I will discuss how hierarchical teams of agents coordinate quantum chemistry workflows and how molecular structure generation addresses a critical gap between chemical reasoning and computational execution. I will consider how these capabilities could contribute to longer-term research goals, such as automated reaction pathway exploration. These examples illustrate how a single missing capability can interrupt an otherwise sophisticated research workflow.
\r\nI will also discuss the open challenge of enabling agents to learn continually from scientific literature, computational results, and failed attempts, and to acquire capabilities as new problems demand them. By examining current capabilities and remaining bottlenecks, I will outline a path toward AI systems that formulate and test chemical hypotheses with increasing independence.