Claude Code will watch a file using tail -f and notice any change to the file waking up. Read cache tokens are 0.1x (10%) the cost of normal tokens so we want to use them. Every time a sub agent starts it has a fresh cache losing that benefit.

Start a few chats with model and effort. Create a folder not tracked by git and put a .md document for each chat. Let the chat know to watch its own file for changes. Then start an orchestrator telling it which chats are watching which files. It will coordinate work between the chats like starting a sub agent but each will have that cheap read cache already set.