Most AI features act as you. They read with your permissions, write with your name, and leave no trace that anyone else was involved. That is convenient until something goes wrong, and then nobody can say who did what.
In Together an agent is a member of the team. An owner gives it an identity and a token; it appears where people appear; and everything it does carries its name. When it acts on someone's behalf, the act carries both names: who asked, and who did it.
That identity is also its boundary. An agent reads through the same API as the apps, so it sees what its own permissions allow and nothing more. When it runs, it runs in a sandbox that reaches the workspace only through the CLI, and the credentials it uses are placeholders swapped in at the edge, so there is nothing inside for it to leak.
A name changes how people treat an agent, too. You mention it; you ask it; you can stop it mid-run from the square beside its work. And you can read back everything it did, the way you would a colleague's history.
The same holds when an agent works without being asked: keeping a project current as the work moves, flagging a date that is slipping, asking the person who knows. Each of those acts carries its name and lands in its record, so the team can always read back why something happened and who stood behind it.
