The Brief
Everything starts with a single sentence: "I need a launch campaign for an AI productivity app." An agentic system parses this into structured intent — audience, tone, deliverables, channels — before any pixels are generated.
Campaign parameters
Changing parameters re-shapes the pipeline: variant count scales the particle burst at stage 6, and channels add distribution spokes at stage 7.
How agentic creative pipelines actually work
Decomposition, not one giant prompt
An agent breaks "make me a campaign" into a task graph: strategy → copy → visual concepting → storyboard → generation → variation → packaging. Each node calls a different model or tool (LLM for copy, diffusion/video model for frames, TTS for voiceover), then passes structured output downstream — the same planner/executor loop used in coding agents.
Why storyboards before video generation
Video models are expensive per second of output. Generating keyframe stills first (a storyboard) lets the agent — and you — approve composition cheaply, then use image-to-video conditioning so motion stays anchored to approved frames. This is the standard cost-control pattern in tools like Runway Gen-3/Gen-4 workflows.
Variant math
Total assets = hero cuts × aspect ratios × channel variants. A modest 3 × 3 × 4 = 36 deliverables from one brief — the real reason agents beat manual workflows: the marginal cost of variant #36 is near zero.
Where humans stay in the loop
Good agent products insert approval gates after concept and storyboard stages. You're not reviewing 36 videos — you're reviewing one strategy doc and nine frames, then letting the machine fan out.