Next gen agentic architecture: Hands on with Gemini 3.5 & ADK

Key Concepts

• Gemini 3.5 Flash: The latest model iteration focusing on high intelligence, improved coding capabilities, agentic workflows, and multimodal efficiency at a lower cost.
• Gemini Omni: A new multimodal model capable of generating video from various inputs (text, images, video) with high consistency and world knowledge.
• Google Cloud Agent Platform: A comprehensive ecosystem for building, scaling, governing, and optimizing AI agents.
• Agent Development Kit (ADK): A framework for building complex, scalable AI agents.
• Agentic Workflows: The ability of models to autonomously plan, use tools, and execute multi-step tasks.
• Evaluation (Evals): A critical process in the agent lifecycle to ensure model outputs align with organizational standards.
• Anti-Gravity: A development tool/CLI that assists in scaffolding and building agents using ADK.

---

1. Google Cloud Agent Platform

The platform is defined by four pillars: Build, Scale, Govern, and Optimize. While developers often focus on the "Build" phase, the speakers emphasized that "Scale, Govern, and Optimize" are essential for production-grade applications.

• Agent Runtime: Provides a secure environment for deploying agents.
• Agent Registry: A centralized repository to manage and track thousands of agents within an organization.
• Evaluation Framework: Integrated testing to ensure agents meet performance and safety benchmarks, which is crucial as models and requirements evolve.

2. Gemini 3.5 Flash: Performance and Capabilities

Gemini 3.5 Flash is designed to provide near state-of-the-art results at a significantly reduced cost.

• Key Improvements:
  • Coding: Enhanced performance on industry-standard benchmarks like SWE-bench.
  • Agentic Workflows: Superior tool-calling capabilities and multi-step reasoning.
  • Multimodality: Native multimodal training allows for better context retention.
• Efficiency: The model offers high tokens-per-second throughput, making it ideal for real-time agentic interactions.

3. Gemini Omni: Generative Media

Gemini Omni represents a leap in generative media, specifically for video creation.

• Capabilities:
  • Consistency: Maintains character and object consistency across video frames.
  • World Knowledge: Understands physics and real-world concepts (e.g., rendering a "claymation" style video of Newton’s First Law).
  • Iterative Editing: Users can modify generated videos using natural language prompts (e.g., adding a hat to a character in a video).
• Real-World Application: The model can take diverse inputs—such as a storyboard image or a reference video—to generate high-quality, 10-second video clips.

4. Step-by-Step: Building an Agent with Anti-Gravity

Lavi demonstrated a workflow for building a "Daily News Bot" using the Agent Development Kit (ADK) and the Anti-Gravity CLI:

1. Define Intent: Provide a natural language prompt (e.g., "Build me a daily news bot using EDK").
2. Scaffolding: The Anti-Gravity CLI automatically generates the project structure (app folder, test folder, etc.).
3. Tool Selection: The model identifies necessary tools and MCPs (Model Context Protocols) to fetch RSS feeds.
4. Implementation Plan: The system creates a task list and implementation plan, which can be reviewed and commented on by human developers.
5. Automated Testing: The agent incorporates built-in evaluation tests to ensure the code meets requirements before deployment.
6. Deployment: Once the logic is verified, the agent can be deployed directly via the CLI.

5. Notable Quotes

• "Gemini 3.5 achieves higher intelligence while bringing down the cost of the frontier." — Dave Elliott
• "You can build agents, you can play with them, but unless they're evaluated and the organization sort of agrees and aligns with that, there's no point of doing that." — Lavi
• "The real unlock of this model is you're able to do iterative video editing through natural language." — Katie Winn

6. Synthesis and Conclusion

The presentation highlights a shift toward agentic development, where the barrier to entry is lowered by AI-assisted coding tools like Anti-Gravity and the high-performance, cost-effective Gemini 3.5 Flash model. Simultaneously, Gemini Omni is democratizing high-quality video production by allowing users to generate and edit complex media through simple natural language prompts. The core takeaway for developers is to move beyond simple prompting and start experimenting with the full Agent Platform ecosystem to build, govern, and scale production-ready AI agents.

Actionable Advice:

• Visit ADK.dev to start building with the Agent Development Kit.
• Use the Agent Platform documentation to understand how to integrate Gemini 3.5 into existing workflows.
• Experiment with Gemini Omni via the Gemini app or YouTube Shorts to understand its creative potential before it hits the API.