Build an AI Video Workflow with an Agent and ArtArch CLI

An AI agent video workflow needs more than a single prompt. It needs a project, connected inputs, visible scene decisions, and a way to wait for the final artifact. ArtArch CLI gives an agent command-line access to those parts of an ArtArch Studio canvas, so a request can grow from one reference-led shot into a repeatable video workflow.

Treat the canvas as the production brief

Start by describing the shot in user-visible terms: the subject, the reference image or video, the setting, the camera movement, the action, and the output format. A brief such as “use this subject in the same room, begin with a medium shot, and move the camera toward the product” gives the agent concrete choices to map onto a video canvas.

Before editing, inspect the current project and its node graph:

artarch project list | jq
artarch project get <canvas-id>
artarch node list

The project view shows the canvas structure, node membership, and connections. That readback helps the agent preserve an existing setup when you are refining a shot or extending a sequence.

Build the shot from connected inputs

Use artarch meta list to confirm the available video node types and their configuration fields. Configure the node for the inputs the shot actually needs, then connect the source image, video reference, or upstream result through the canvas.

ArtArch CLI checks the connection and persists the target input slot. For list inputs, the target handle selects the correct position. A completed generation with a persisted CDN result can also be reused as an input, which lets one approved shot feed the next step in a longer sequence.

The agent should keep one visible beat and one main camera move in each short video node. That makes a change easy to describe: adjust the subject action, change the camera move, or replace the reference while keeping the rest of the canvas readable.

Run the workflow and keep the run ID

When the graph is ready, review the full scope. A flow run can submit the whole canvas and may consume credits. Then start the existing workflow and wait for its terminal state:

artarch flow run --wait --wait-timeout 30m

The CLI prints the result, including artifact URLs when the run completes. You can download the artifacts immediately with --download, or fetch them later with:

artarch run download <run-id> --node <node-id> --output ./results

If a wait reaches its timeout, continue with artarch run wait <run-id>. Keep the same run ID. Resubmitting the flow would create a second generation request and make the project's state harder to reason about.

Video scenes that benefit from an agent

A creator planning a short product scene can ask the agent to connect a product image, define one camera move, run the video node, and download the take for review.