Publish once, update every workspace

Experts change their minds in small ways. A question gets sharper. A summary format loses a section nobody read. A warning gets added after a near miss. The work is not writing the insight down once. The work is making the next engagement use the new version without a scavenger hunt.

Copying an agent per client feels safe because the copy cannot affect the others. It also cannot receive a fix. Six months later you have six instructions, four of them missing the warning, and no record of which one is current. The clients are isolated, and so is your improvement.

Publishing is the other pattern. The agent has a draft you can break. Publishing freezes a version: the instructions, the skills, and a changelog, written by someone, at a time. Linked spaces point at the published version. The next publish moves that pointer. Rollback points it back. The draft never leaks out just because you typed in it.

Client data does not ride along. A version is a snapshot of the agent, not of the space. Memory notes, uploads, and conversations stay in the room where they were created. That is the whole point of linking instead of cloning. You improve the shared part once. You do not gather everyone's files into the place you edit.

This release stores that shape and does not yet let you drive it from the builder. You can create the organization and the client space. You cannot publish an agent from the interface. The pages say so. The model is already there so the next phase does not have to invent a second notion of "current."

When you do publish, write the changelog as if a colleague will ask why the agent's tone changed. They will, and the audit log is a better answer than a message you cannot find.