What's this about?
Agents increasingly do knowledge work — producing and modifying persistent artifacts like code repositories, documents, spreadsheets, slides and reports. But the parsed views they work through don't handle versioning well. StagedWorkspace proposes a versioned workspace designed for that.
What exactly is a "parsed view"?
An agent doesn't see your spreadsheet the way you do. Something converts the file into a representation the model can read — a text rendering of the cells, an outline of the document. That parsed view is lossy and, more importantly, usually stateless: the agent reads, decides, writes back, and nothing tracks how the artifact evolved or which version the decision was based on. When several agents work the same artifacts, or when one runs long enough that things change underneath it, that becomes a correctness problem rather than an inconvenience.
Why is this worth noticing?
- It's the same insight as Agentic Transaction from last week, one layer up: agents that modify persistent state need the guarantees we long ago built for databases and source control.
- Knowledge work is where agent deployment is growing fastest outside coding, and code already has version control — documents and spreadsheets don't.