File Systems Are the New Primitive for AI Agents

A growing consensus among AI researchers and developers holds that file systems—not databases or vector stores—are becoming the fundamental primitive for building capable AI agents. Rather than relying on complex, opaque internal databases, these new agents read, write, and organize files using standard operating system interfaces, much like a human user would.

Proponents argue this approach offers several advantages. File systems are inherently persistent, universal, and human-readable. An agent that saves its state to a file can be inspected, debugged, and resumed by a human or another agent. Folders become natural scopes for context and memory; filenames and directory structures provide built-in search and organization without requiring a separate indexing layer.

The idea has been bubbling up through projects like Auto-GPT and BabyAGI, which already use file system operations for task planning and memory. But the thesis goes further: the file system itself should be the primary interface for an agent's internal and external operations, replacing bespoke task queues and memory stores with standard files, symlinks, and directories.

If this vision holds, it could make AI agents more transparent—users could literally open a folder and see what their agent is working on—and more interoperable, since every operating system already knows how to handle files. Rather than building new infrastructure for agents, developers can leverage decades of mature operating system design.