Google’s new AI agent stack from I/O 2026

Key Concepts

• Agent Stack: A four-layered framework for building, orchestrating, and deploying AI agents.
• Gemini 3.5 Flash: A new, high-speed, cost-effective model optimized for long-horizon agentic tasks.
• Agent-to-Agent Protocol: A universal communication standard allowing different agent solutions to interoperate.
• Skills Registry: A private, organization-scoped repository for reusable agent tools and instructions.
• Anti-Gravity: An agent-first development platform for orchestrating multiple agents.
• Gemini Omni: A natively multimodal creation model capable of reasoning across text, images, audio, and video to generate coherent video content.
• SynthID: A watermarking technology used to identify AI-generated content.

---

1. The Agent Stack: A Four-Rung Ladder

The speaker outlines a structured approach to moving agents from development to production, emphasizing that developers can start at any level and scale up using the same underlying platform and protocol.

• Rung 1: Agent Studio (Low-Code): Designed for users who prefer a visual UI to build agents, with the ability to export to code for advanced development.
• Rung 2: Managed Agents API: A fully hosted, configuration-first approach. It eliminates the need for manual infrastructure management (sandboxes/environments) and supports persistent state across sessions.
• Rung 3: Anti-Gravity: A platform for managing and orchestrating cohorts of agents. It includes a desktop app ("Mission Control") for steering parallel tasks, dynamic sub-agents, and scheduled automation.
• Rung 4: SDK 2.0 (Code-First): Provides full control for experienced engineers. It has transitioned from imperative to graph-based orchestration, allowing developers to define workflows as a graph of steps. This enables a "slider" approach to balance between deterministic and creative outputs.

2. Development Tools and Infrastructure

• Agent CLI: A unified command-line interface that simplifies the end-to-end lifecycle: scaffolding, local testing, writing evaluations, and deploying to the Agent Platform.
• Agent Platform: An autoscaling, serverless infrastructure that allows agents to run within a secure boundary, inheriting organizational data protections.
• Agent Payment Protocol (AP2): A guardrail layer for commerce-related tasks. It allows users to set "mandates" (spending limits and rules) within which an agent can plan and execute transactions.

3. Skills Registry: Dynamic Governance

The Skills Registry acts as a centralized library for organizational knowledge.

• Mechanism: Skills are defined as markdown files containing instructions and tools.
• Dynamic Discovery: Agents do not carry all skills; they query the registry at runtime to load only the specific skills required for the current task.
• Benefits: Enables shared governance, reduces agent "bloat," and ensures all agents in an organization have access to the latest approved tools.

4. Gemini 3.5 Flash and Gemini Spark

• Gemini 3.5 Flash: Co-designed with Google’s TPUs, this model is optimized for "long-horizon" work (tasks running for minutes or hours). It outperforms previous Pro models on agent-specific benchmarks.
• Gemini Spark: A 24/7 personal agent that runs in isolated, ephemeral VMs. It learns user context (writing style, collaborators) and executes multi-step tasks across Google Workspace while maintaining security and privacy.

5. Gemini Omni: Multimodal Creation

Gemini Omni represents a leap in generative AI, moving from simple generation to "creation through reasoning."

• Functionality: Unlike tools that rely on a single prompt, Omni fuses multiple inputs (images, sketches, voice clips, text, existing video) into a coherent scene.
• Grounded Reasoning: It understands physics, gravity, and momentum, allowing it to manipulate elements in a video (e.g., making a mirror ripple like liquid) while maintaining consistency in lighting and character identity.
• Real-World Application: The speaker demonstrated using Omni to synthesize disparate photos of themselves and event footage into a single, coherent video clip.
• Safety: Every video generated by Omni includes a SynthID watermark to ensure transparency regarding its AI-generated nature.

---

Synthesis and Conclusion

The presentation highlights a shift toward a unified, scalable ecosystem for AI agents. By providing a consistent "ladder" of tools—from low-code visual interfaces to graph-based SDKs—Google aims to lower the barrier to entry for production-grade AI. The integration of the Skills Registry and the Agent-to-Agent protocol addresses critical enterprise needs for governance and interoperability. Simultaneously, the launch of Gemini 3.5 Flash and Gemini Omni demonstrates a focus on high-performance, long-horizon reasoning and advanced multimodal creation, providing developers and users with powerful primitives to build complex, secure, and context-aware AI systems.