Comment by jpollock

I'm wondering if in large projects, you want subagents to avoid having tasks flush out the main context?

If you're working with large source files, you might want to do each piece of work in an independent context with the information discarded afterwards?

Is the context a sliding window, or are there tiers of importance?