Uploaded November 2026 | Updated November 2026, 2 hours ago
Most AI agents fail in production because a single prompt can’t guarantee order, safety checks, or human approval. Today, we’re building a fully autonomous, multimodal AI video studio from scratch using Google’s Agent Development Kit.
Imagine running a daily cozy-spooky video channel ("VibeTube"). Every night, your audience sends you chaotic ideas: voice notes, porch videos, spooky photos, and text comments. Gemini can read them, Imagen can draw the posters, and Veo can render the video.
The problem? Generative models can do the chores, but they can’t run the studio. They can't refuse a bad idea before it burns your API budget, pause for your approval, or wait for a slow video render without timing out. A prompt can’t promise order. A workflow graph can.
In this live build, we are walking through the VibeStudio (Haunted Edition), architecture to turn a chaotic creative process into a deterministic, production ready ADK Workflow on Google Cloud.
*What the team is building live:*
* 🧠 Beyond basic prompts: Moving from single turn agents to structured ADK Workflow graphs (fan-out, parallel JoinNodes, and smart routers).
* 👁️ Multimodal RAG ingestion: Using Gemini to transcribe voice notes, caption photos, and watch video clips—feeding them all into a single Vertex AI RAG Engine corpus.
* 🛑 The 2 types of suspension: How to make your agent pause for a human-in-the-loop decision and asynchronously wait on long running Veo video renders without holding connections open.
* 💾 Agent memory: Wiring up Memory Bank callbacks so your studio actually remembers what it did yesterday.
* 🚀 Prompt to production: Deploying the entire VibeStudio workbench to Cloud Run and publishing our first AI generated short.
Resources:
👉 Follow along with the Codelab → https://g.dev/ai/vibe-studio
💻 GitHub repo → https://g.dev/ai/vibe-studio-multimodal
Speakers: Annie Wang, Logan Hennessy
Products Mentioned: Agent Development Kit, Veo
Most AI agents fail in production because a single prompt can’t guarantee order, safety checks, or human approval. Today, we’re building a fully autonomous, multimodal AI video studio from scratch using Google’s Agent Development Kit.
Imagine running a daily cozy-spooky video channel ("VibeTube"). Every night, your audience sends you chaotic ideas: voice notes, porch videos, spooky photos, and text comments. Gemini can read them, Imagen can draw the posters, and Veo can render the video.
The problem? Generative models can do the chores, but they can’t run the studio. They can't refuse a bad idea before it burns your API budget, pause for your approval, or wait for a slow video render without timing out. A prompt can’t promise order. A workflow graph can.
In this live build, we are walking through the VibeStudio (Haunted Edition), architecture to turn a chaotic creative process into a deterministic, production ready ADK Workflow on Google Cloud.
*What the team is building live:*
* 🧠 Beyond basic prompts: Moving from single turn agents to structured ADK Workflow graphs (fan-out, parallel JoinNodes, and smart routers).
* 👁️ Multimodal RAG ingestion: Using Gemini to transcribe voice notes, caption photos, and watch video clips—feeding them all into a single Vertex AI RAG Engine corpus.
* 🛑 The 2 types of suspension: How to make your agent pause for a human-in-the-loop decision and asynchronously wait on long running Veo video renders without holding connections open.
* 💾 Agent memory: Wiring up Memory Bank callbacks so your studio actually remembers what it did yesterday.
* 🚀 Prompt to production: Deploying the entire VibeStudio workbench to Cloud Run and publishing our first AI generated short.
Resources:
👉 Follow along with the Codelab → https://g.dev/ai/vibe-studio
💻 GitHub repo → https://g.dev/ai/vibe-studio-multimodal
Speakers: Annie Wang, Logan Hennessy
Products Mentioned: Agent Development Kit, Veo










