good pattern from this google paper the verifier starts from a clean context, so the explorer's reasoning can't talk it into a bad proof so the explorer can try wild ideas and the verifier tosses the wrong ones same reason deepagents subagents get their own context

→ View original post on X — @hwchase17