Idea replies — I built a control room that runs many AI coding agents for my studio. What else could it become?

Good question, it's the part I spent the most time on. The agents run on my own Mac under my user, so in theory they can reach a lot. What keeps it sane: credentials sit in a local vault and get pulled one at a time only when a task needs them, never pasted into prompts. Anything that moves money, deletes for good or touches someone else's data stays frozen until I approve it, and payment accounts that belong to clients are read-only, always. Each session also declares on a shared board which files it's touching, so two agents don't step on each other. Honestly the bigger risk I found wasn't a rogue agent, it was a confident one, so a lot of the rules are about proving the work before calling it done.