I recently asked on Reddit whether filesystems make sense as an interface for AI agents. Models are already familiar with paths and commands like ls, cat, and grep, so giving them files to work with seemed like a natural starting point.
The discussion led into what happens after an agent opens a file: another agent edits it, the workspace grows, or a later run needs to undo a mistake. These are the three problems I came away thinking about.
Two agents can overwrite each other's work
Suppose two agents read the same file. One updates it and saves. The other then saves an edit based on the older version. Without checking for changes in between, the second save can overwrite the first agent's work.
Both writes can succeed even though work has been lost. To prevent that, the workflow needs to catch stale edits before accepting them and provide a way to reconcile the changes.
Finding the right context takes more than file access
An agent may have access to thousands of files without knowing which ones matter. It can end up listing folders, searching for keywords, opening files, and trying again when the results aren't useful. Search results and file contents brought into its context consume tokens, including material that never helps answer the question.
Repeating that process across a task can make finding information expensive before the agent does anything useful with it. An index or targeted retrieval tool can help narrow the search and bring back only the relevant material.
History gives you a way to review and recover
An agent can misunderstand an instruction, remove useful information, or make an incorrect edit. Once that change is saved, another agent may build on it before anyone notices the mistake.
Keeping a history lets you review what changed and recover when something goes wrong. A diff shows the edits, a record of who made them helps trace the problem, and an earlier version gives you something to restore.
How this connects to AgentWS
Git, persistent folders, and existing search tools can support useful agent workflows today.
Reviewing a few changes from one agent is manageable. Keeping up with several agents editing shared files throughout the day is a different challenge. When changes arrive faster than people can inspect them, workflows that depend on manual review and conflict resolution become harder to sustain.
That's why I'm building AgentWS, a persistent filesystem and workspace with a control plane designed for agents. The goal is to manage the workspace throughout an agent's work: bringing in application data, controlling access, coordinating changes across agents, and preserving results so work can be reviewed, recovered, and continued across runs.
Follow my journey on X as I build AgentWS, test ideas, and share what I learn along the way.