We are seeing an explosion in the use of autonomous agents in developing software. Such agents, often constructed using language models, now perform tasks from issuing queries to explore data stored in databases, to compiling and generating test code. In this talk, I will first share our observations in studying agentic workflows for solving the natural language to SQL task, and outline our vision in building data systems to efficiently execute such workflows. I will then describe our current efforts towards that vision, specifically our work in using agents to optimize Pandas workflows, and to generate code to execute on various hardware accelerators. I will conclude by describing a new benchmark we are creating to measure the progress of agents in solving coding problems.