Agent Architecture as Folders

One of the more interesting ideas floating around the AI agent space right now: use the filesystem as your architecture. Agents as directories. Memory as files. Coordination through the thing that's already there.

It sounds like a simplification to the point of absurdity until you think about why most agent frameworks are complicated, and what that complexity is actually buying you.

Why frameworks get heavy

Multi-agent systems pick up complexity fast. State needs to live somewhere. Agents need to communicate. Work needs to be divided and reassembled. Frameworks try to solve all of this with custom infrastructure — message queues, shared memory abstractions, orchestration layers.

The folder approach says: you already have all of this. Files are state. Directories are namespaces. File watching is event-driven coordination. The operating system has had these primitives for decades.

The tradeoff

What you lose is type safety and explicit contracts between agents. A file can contain anything; a message queue entry has a schema. For small, controlled systems this doesn't matter. For anything that needs to scale or be maintained by a team, you'll eventually want more structure.

But as a starting point — as a way to prototype agent systems without buying into a framework's opinions — this is legitimately underrated. Some of the most maintainable code I've seen starts with the simplest possible representation of the problem.

Folders are a good representation of a lot of problems.