Yeah, that sounds right. So when you do the first scenario (i.e. running outside of a Git repository), ClearML will just take the code file content and embed it on the task - that's why the agent can take it an run it (even if it doesn't have access to the git repository, because it's not required).
In the second scenario, ClearML basically stored the git repository details on the task, and the agent needs to access it somehow (using some sort of authentication)