I've also gone this route. A little board in tmux I pull up with a key combination, some indicators of activity / needs me to take a look at it and some background machinery / limitations as my work is systematically: open a worktree, define separate correctness and performance tests, run the latter on a chosen cluster, and then eventually open a PR. Testing is handled via another vibecoded regression tool/databse. The agents mostly use the board commands for syncing, rebasing & pushing, etc. To avoid errors. The PR hierarchy is encoded in it. Very convenient for me, probably useless to most!