Running several command-line coding agents at once can save time, but only if you can see what each one is doing. The failure mode is not the model; it is losing track of which agent is waiting for approval, which one is stuck, and which one already changed a file.
If every agent has its own terminal, you are the queue. A shared task list with one approval gate keeps the work visible and prevents two agents from editing the same file at the same time. An agent that can run any command without review is faster until it is not. A single approval point per task is easier to audit than a separate confirmation for every tool call.
Task history should stay on your machine. If the collaboration layer runs locally and only relays traffic, you keep the logs and the working files; the remote service does not need them.
The useful test is simple: start two agents on two small tasks, then see whether you can tell which one needs you without opening a second window. If yes, the layout is doing its job.