“Let the AI advance the story” sounds exciting, but without a state model, every chapter quickly starts inventing the same problems again.
Cogweald models long-running plot lines as story threads. Each thread has a title, description, status, and timestamps. Its status can be active, resolved, or abandoned.
When a chapter is generated, the model does not only return prose. It also returns thread updates: which thread advanced, which one was resolved, and whether a new one should be opened. The application applies those decisions to the database and provides the updated state to the model next time.
This turns continuity from a hope into an observable and verifiable state transition.
The same pattern works for any agent that needs long-term planning:
- An unfinished refactor for a coding agent
- An unverified hypothesis for a research agent
- A customer risk for a success agent
- A quest line for a game agent
If an agent needs to remember what is still unfinished across multiple runs, it probably deserves a state machine.
Top comments (0)