Automating Creativity: Building Gen Media Agents with ADK and MCP
Summary
In this live-coding session from Google Cloud Next, Developer Relations Engineer Katie Nguyen joins host Stephanie Wong to demonstrate how to automate generative media workflows using Google's Agent Development Kit (ADK) and Model Context Protocol (MCP) servers. The session serves as a programmatic companion to a prior session by Khulan, who focused on manual creative prompting techniques.
Katie walks through a working 'Character Story Agent' she built using ADK, which orchestrates multiple Google Gen Media models to take a simple text description and produce a fully narrated, multi-scene short video — entirely through natural language interaction. The live demo centers on Lulu, a Shih Tzu, who gets into mischief while home alone. With minimal user input (just the breed and name), the agent autonomously generates a three-scene story arc, produces character-consistent images, animates them into 6-second video clips, composes background music, and layers in expressive narration.
The agent's tool stack is connected via MCP servers and includes: Nano Banana 2 for high-fidelity, prompt-adherent image generation; Veo 3.1 Light for fast image-to-video animation; Lyria 3 Pro for mood-matched musical scores; and Gemini 3.1 Flash TTS for expressive narration with audio tags. Open-source tools like FFmpeg are invoked via command line for post-production — stitching audio, video, and music without human editing.
Katie explains that the key benefit of the agentic approach over manual prompting is consistency: the agent maintains a 'memory bank' of character details (Lulu's bow, fur texture, environment) and uses the initial reference image as a baseline for all subsequent scene generation. This ensures character continuity across images, videos, and audio that would be tedious to manage manually.
The session also covers Agent Skills — modular instruction sets that can be loaded into the agent to offload complex logic from the system prompt. Katie demonstrates a 'Voice Director' skill (from the open-source Gen Media MCP GitHub repo) that provides the agent with detailed knowledge about expressive audio tags, voice profiles, and best practices for getting the most out of Gemini TTS. A parallel 'Image Artist' skill does the same for Nano Banana prompting.
On evaluation, Katie describes a pattern where a separate 'image evaluator agent' can act as an LLM-as-judge: checking whether generated assets adhere to the original prompt, verifying character consistency (e.g., is Lulu's bow in the right place?), and detecting technical issues like audio running longer than the video clip — triggering automatic regeneration. This self-correction capability reduces the need for constant human oversight in the creative loop.
The session closes with a live demonstration of natural language post-production: Katie asks the agent to make the background music louder in the final video, and the agent handles the audio adjustment and re-renders the file autonomously. Both the Gen Media MCP server repository and ADK Code Labs are shared for developers to get started.
Key Takeaways
- •Agentic workflows maintain character consistency across images, video, and audio by using an initial reference image as a memory anchor — eliminating the manual effort of re-prompting with character details each time.
- •MCP servers let ADK agents connect Gemini to specialized Gen Media tools (Nano Banana, Veo, Lyria, TTS) without custom API wiring for each model.
- •Agent Skills decouple complex prompting logic from the main system prompt — load a 'Voice Director' or 'Image Artist' skill when needed rather than bloating the instruction context.
- •Open-source tools like FFmpeg can be integrated as command-line tools within an agentic framework to automate video stitching and audio mixing without a human editor.
- •A separate evaluator agent acting as LLM-as-judge can fact-check generated assets against the original prompt and trigger automatic regeneration when outputs fall short (e.g., audio overrunning video length).
- •Gemini's multimodal capabilities enable image and video evaluation — not just text — allowing agents to verify visual consistency like character details across scenes.
- •Natural language post-production is viable: users can ask the agent to 'make the music louder' or 'regenerate scene two' and the agent handles the technical execution autonomously.
- •Keeping videos short (6 seconds per scene) during prototyping reduces generation time and cost while still demonstrating the full pipeline end-to-end.
- •The Gen Media MCP GitHub repository provides ready-to-use MCP servers, sample agents, and skills (Voice Director, Image Artist) that developers can drop into their own projects.
- •Engineers without creative backgrounds can use Gemini as a creative collaborator to generate story ideas, scene descriptions, and prompt details — solving the blank-slate problem.
Transcript
Everyone and welcome back. We are here at now with Katie Wynn who is a developer relations engineer here at Google Cloud. Uh, welcome to the show Katie first of all. >> Yeah, thank you so much for having me. I'm so excited to be here. >> Yeah, well we just wrapped up a conversation with Kulon about Gen Media. She's the creative crushing it with the prompts, but how can we actually maybe automate some of the workflows and that's what you're going to talk about, right? >> Yeah, absolutely. So I'm ...
