ai video looks pretty damn good now. but getting the exact shot you want is still a pain.
you generate a perfect scene, except the camera angle is wrong.
try fixing it with another prompt, and suddenly the objects, lighting, and entire composition change. now you're generating everything again.
intangible 2.0's new Canvas lets you build an actual 3D scene underneath the footage.
you can move the camera, reposition objects, and generate another shot while keeping the underlying scene intact.
that's the kind of creative control ai filmmaking has been mission.
i spent two days in the DSH super agent source. if you’re building agents, here’s what’s worth taking from it:
1. keep routing and execution separate. one router talks to people and assigns jobs. each job gets a worker with its own context.
2. filter events in code. check the source, sender, and permissions before calling a model. run proactive rules in observe mode first so you can see what they’d act on.
3. give the job a ticket. record what authorized the work and how far it can go. replay the same trigger and you still get one ticket
4. send follow-ups straight to the worker. pr comments, approvals, and ci results can move the job forward without another router call.
5. build for crashes. save outgoing replies before sending them. a retry can find the github comment it already posted, and a restarted worker can pick up in the same workspace.
what you get from building this way: you stop babysitting 3 to 5 sessions, work survives crashes and restarts, token spend drops because most events never reach a model, and a human only gets pinged when a decision is needed.
@jino_rohit You have to try pi with codex /claude & ponytail, and make some local evals & judges accordingly to your need and add custom tool calls in pi crazyy experience so for