Google Cloud Next '26• Apr 23, 2026• 20:03ConferencePanel

Automating Creativity: Building Gen Media Agents with ADK and MCP

From Google Cloud Next '26 · 2026

Katie Nguyen•Developer Relations Engineer, Google Cloud

Executive Summary

Google Cloud's Agent Development Kit (ADK) enables the creation of agentic workflows to automate complex, multi-step generative media tasks, from story ideation to final video production.
The framework orchestrates a suite of specialized models, including Gemini for logic and memory, Nano Banana 2 for image generation, Vio for image-to-video animation, and Lyria for music, to create cohesive stories.
Developers can use reusable, open-source 'agentic skills' (e.g., 'Gen Media Voice Director') to provide agents with specialized domain knowledge, improving the quality and control of outputs.
The system supports automated evaluation, where a separate 'evaluator agent' can use an LLM to judge generated content for quality, consistency, and adherence to prompts, enabling self-correction.

12 quotes

Concerns Raised

The generated media can have synchronization issues, such as narration running longer than the corresponding video clip, which requires either manual iteration or a more advanced self-correction loop.
While the end-user interaction is simple, the initial setup of the agent, its tools, and skills requires significant engineering effort and understanding of the underlying SDKs and APIs.

Opportunities Identified

Automating the entire creative pipeline from a simple concept to a finished video can dramatically accelerate content creation for marketing, entertainment, and education.
The use of natural language to interact with and iterate on complex media projects lowers the barrier to entry for individuals without deep technical or creative prompting skills.
The modular 'skills' framework allows for the creation of a marketplace or ecosystem of specialized agent capabilities, enhancing the power of the platform.

Key Themes

Agentic Media Generation

The core concept is using an AI agent to automate the entire creative pipeline. A simple natural language prompt about a character triggers a complex sequence of tasks: storyboarding, image generation, video animation, narration, and music composition, all without manual intervention.

This approach dramatically scales content creation, reduces manual labor, and allows creators to focus on high-level concepts rather than the tedious, step-by-step process of media production.

Multimodal AI Orchestration

The agent acts as a conductor for a diverse orchestra of specialized AI models. It leverages different models for text (Gemini), images (Nano Banana 2), video (Vio), and audio (Lyria, Gemini TTS), seamlessly integrating their outputs into a single, cohesive final product.

Complex creative tasks require a combination of specialized AI capabilities. This demonstrates a practical framework for orchestrating multiple models to achieve a goal that no single model could accomplish alone.

Ensuring Character and Story Consistency

A key challenge in generative media is maintaining consistency. The agent addresses this by first generating a baseline character image and then using that image as a reference for all subsequent scenes, ensuring the character's appearance remains stable throughout the story.

This technique is crucial for creating believable narratives and maintaining brand identity in automated content. It solves a common failure point in generative storytelling.

Modular and Reusable AI Skills

The Agent Development Kit (ADK) allows developers to encapsulate domain knowledge into modular, reusable 'skills'. For example, a 'Voice Director' skill instructs the agent on how to best use text-to-speech models, including nuances like emotional tags and voice selection.

This modular approach makes AI agents more powerful, scalable, and easier to maintain. It allows for the separation of logic, enabling teams to build and share specialized components for their agents.

Automated Evaluation and Self-Correction

The framework supports creating a separate 'evaluator agent' that acts as a quality control mechanism. This agent can use an LLM to analyze generated media, check it against the original prompts for consistency and accuracy, and identify errors like audio-video desynchronization for correction.

This moves towards more autonomous AI systems. Building self-correction loops reduces the need for constant human oversight and improves the reliability and quality of automated workflows.

Get started free

Topics

Generative AI Agentic Workflows Google Cloud Gemini Multimodal AI Text-to-Image Image-to-Video Text-to-Speech AI Music Generation Agent Development Kit (ADK)Nano Banana 2 Vio Lyria AI in Media Content Automation LLM as a Judge Character Consistency

Processed Apr 28, 2026 yt-dlp + mlx-whisper + Gemini