We've been trying to put LangGraph agents into production for a while. The thing that kept biting us was tool-call boundary enforcement: stuff like "must call X before Y", "max N retries", "approval gate before destructive action". Worked fine in demos, broke at the moments that mattered. What we tr
