When an AI job runs once a day, it is easy to underestimate the distributed-systems problems hiding underneath.
What if the scheduler retries? What if two workers claim the same world? What if the model succeeds but the database connection drops before the chapter is stored?
Cogweald treats daily advancement as a claim-based transaction flow. A world’s last_tick_at acts as both the schedule marker and the claim token. A worker must successfully claim the tick before it is allowed to generate a chapter.
That small design prevents one of the most dangerous failure modes: cron jitter producing two chapters for the same world on the same day.
Reliability in AI systems often comes less from a better prompt and more from ordinary-looking conditional database updates.
Once an agent can change the world on its own, design its state transitions with the same care you would use for a payment flow.
Top comments (0)