Long Work: Chaining and Agents

A million-token context window does not make a million-token conversation a good idea. Long work fails in four consistent ways: settled decisions get reopened, early constraints fade, scope grows, and unverified work gets reported as complete - and a bigger window makes all four *more* likely, because it removes the limit that used to force you to split the job. This book is the structure that prevents them: stages split at the seams where the thinking changes, handoffs written rather than scrolled, an explicit state block with a fenced list of decisions, scope held against helpful drift, verification required instead of assumed, and gates set by reversibility.