Real work rarely needs one agent. A merge request review might want a reviewer, a tester and a security check, each with its own instructions and model. Grouping them as a team gives that set a name, a note on when to use it, and one place to change it.
The hard question is what happens when one member changes. Some teams must not change underneath the people relying on them: they were tested with a specific text. Others should always get the newest version. A team that records the choice per member makes the difference explicit instead of accidental.