Human-agent work has a weak point: the agent often produces the final artifact and disappears. The person receives the result as a “finished file”, not as a surface where both sides can keep thinking together.
The shared document pattern changes that. The person edits by hand, the agent edits through code, but both operate on the same visible object. This is closer to a document, canvas, or working room than to a chat with an attachment.
The connection with tldraw and the “one document, two hands” idea points to a larger direction: UI becomes a shared state surface. The agent does not need to narrate every diff in text if the changes are visible in the shared artifact.
Editorial read: the next good agent UI is not a “prompt -> result” form. It is a place where intent, manual edits, and programmatic edits remain inside one document.