That's indeed very hard - we're building runtime captures and they are fed to the coding agent (claude code, cursor, etc) to update the agents' codes. However, the runtime data needs to be structured in a certain way so the coding agents won't get confused. We're testing this a lot right now.