The exponential growth of the Internet has produced a labyrinth of databases and services. Almost any type of information is available somewhere, but most users can’t find it, and even experts waste untold time searching for appropriate information sources. Many researchers argue that software agents will remedy the situation by acting as personal assistants and automatically accessing multiple sources, integrating informa-tion, and acting on the user’s behalf. Indeed, this vision is so widely accepted (at least in the AI community) that I won’t dwell on it further. Instead, I’ll describe some of the software robots we've built at the University of Washington, explain why I think planning is a crucial technology for their control, summarize the lessons we've learned by their construction, and suggest directions for future work in the area. Note that this paper is not intended to he a comprehensive survey of important softbot projects -- there are far too many interesting systems for me to describe them all. I focus on University of Washington projects, because I am most familiar with this body of work and because the Washington softbots have emphasized planning-based control.