Picture a pipeline with five steps. An AI agent reads a request, three more research it in parallel, and a human signs off at the end. It runs fine for weeks. Then one night the process dies at step three.

To recover, the system needs to answer two questions. Where was I? Which steps already ran, what did they return, which ones must not run again. And where am I going? What comes after step three, which branch to take, who is still waiting for an approval.