Autonomy is not the only goal
Narrative work contains subjective choices. Two images may be technically valid, but only one fits the beat. Full autonomy often produces an average answer instead of a specific one.
Where an agent adds leverage
The work between judgement and execution is rich with structured tasks.
- Turn story constraints into shot inputs
- Check identity and spatial continuity
- Consolidate findings into one repair plan
- Track asynchronous tasks, versions and failures
Confirm around cost and irreversibility
Let text cleanup and parameter assembly advance automatically. Ask for approval before expensive model calls, destructive changes or final delivery.
Conversation must connect to artifacts
If an agent can only offer advice but cannot reference a shot, update a brief or create a revision, it is another chat box. Its value comes from operating on explicit project objects with traceable actions.