Video as a tool call for agents.

Agents already call search, code interpreters, and browsers. Video should be another tool output, not a separate app the user has to open.

The interface is simple: submit a prompt or upstream model answer, poll job status, receive a video_url. Cheap enough that a tutoring bot can afford to call it.

That is why we care about unit cost and near-real-time latency. If a render costs too much or takes too long, agents will never make video the default.

The next wave of LLM products will not only answer. They will show.