Workspace state is what an agent keeps in its own environment between steps: files, notes, scratch artifacts, a record of what it already tried. It persists across turns, so the agent can look at what happened last time instead of re-deriving from the prompt.
This is the safe side of the write-access boundary. An agent can be messy and wrong here because nothing in its workspace binds anyone else. A system-of-record does bind others, which is why moving a result from workspace into record is the operation that costs trust. The discipline is keeping tentative thinking here and letting only a checked result cross that line.
The fuller argument for why a durable environment changes agent behaviour is in agent-workspaces.