• Inference Systems

Over the past three and a half years, the agent ecosystem has accumulated a growing stack of infrastructure: complex data structured to represent state, which includes an evolution from vector stores to graph based agents to black-box systems. Yet despite all this tooling, many of the most capable long-running agents are converging on something much simpler: a basic filesystem.

In this talk, I will argue that running an agent in a loop over a filesystem is one of the best ways to build generalist agents. It naturally handles state, memory, and intermediate work, letting agents scale from quick tasks to massive workflows without convoluted orchestration. We will look at how agent architectures got here, why the filesystem approach is working so well today, and when sandboxed environments actually help versus when they just add unnecessary complexity.