I use a similar approach for some types of work. I use a planner (read only), researcher (read only), implementer, and reviewer (read only). I allow the implementer to spawn researchers and reviewers to help it with its work. I interact with a controller to manage the subagent trees. The agents use a file-based communication system. I get really good results on real work. Coworkers (using Claude with fable) can’t tell the difference.

The benefit of Claude is purely team collaboration and some basic QoL features.