From the Next ‘26 main stage to the terminal

Key Concepts

• Agent Platform: An end-to-end ecosystem for building, scaling, governing, and optimizing AI agents.
• Agent Development Kit (ADK): A framework supporting Python, Go, TypeScript, and Java for building agents.
• Agentic Data Cloud: A shift from "systems of intelligence" (dashboards/reports) to "systems of action" (agents executing tasks).
• Inference vs. Training: The transition where inference workloads now eclipse training in volume, necessitating specialized hardware like TPUv8.
• Generative Media (Gen Media): A suite of models including Nano Banana (image), Veo (video), Lia (music), and Gemini Audio (speech).
• Vibe Coding: A paradigm where developers use natural language prompts to build, iterate, and deploy applications rapidly.

1. The Gemini Enterprise Agent Platform

The platform addresses the "prototype-to-production" gap. While building a prototype is easy, production requires robust identity, governance, memory, and anomaly detection.

• Governance: Features include a secure gateway, cryptographically generated agent identities, and audit trails.
• Memory Bank: A generally available feature that allows agents to maintain state across sessions, enabling "long-running agents" that can operate autonomously for days or weeks.
• Agent Evaluation: A new pillar focused on non-deterministic testing. Since LLMs are non-deterministic, evaluation dashboards are critical to ensure agents meet business goals.
• Sandboxes: Essential for limiting the "blast radius" of autonomous agents, ensuring they operate within defined permissions.

2. Data Cloud and System of Action

Yasmin Ahmed (Managing Director, Data Cloud) emphasized that data strategy is no longer just about "clean data" but about contextual understanding.

• Knowledge Catalog: Uses GenAI to infer schema, relationships, and descriptions from unstructured data (PDFs, etc.), which agents use to reason accurately.
• Cross-Cloud Lakehouse: Leverages open standards like Apache Iceberg to allow agents to access data across AWS, Azure, and Google Cloud without complex pipelines.
• Efficiency: BigQuery has seen a 35% improvement in processing speed and a 40% cost reduction, while Spark with the "Lightning Engine" is 5x faster than standard Spark.

3. Generative Media (Gen Media)

The Gen Media portfolio allows for high-fidelity creative output with precise control.

• Nano Banana: Used for image generation with specific artistic control (e.g., camera lens, film stock, lighting).
• Veo 3.1 Light: A cost-effective video generation model that supports "first-frame/last-frame" control for precise animation.
• Lia 3 Pro: A music generation model that understands timestamps, allowing users to sync music shifts (e.g., a lullaby or upbeat dance track) to specific video events.
• Gemini 3.1 Flash Live: Features a "Live Avatar" capability that connects to Google Search for real-time, interactive, audio-to-audio conversations.

4. Infrastructure and Inference

• TPUv8: A significant announcement splitting training and inference chips to prevent "traffic jams" in silicon.
• Nvidia Partnership: Google Cloud is integrating Blackwell GPUs (RTX Pro 6000) with 96GB VRAM, allowing multiple models to run on a single GPU.
• Optimization SDKs: Use of TensorRT-LLM and Nvidia Dynamo is recommended to maximize performance on Nvidia hardware.

5. Developer Experience: AI Studio & Vibe Coding

Logan Kilpatrick (AI Studio) highlighted the evolution of "vibe coding"—building apps via natural language.

• Build Tab: Allows users to go from prompt to deployed app (via Cloud Run) in minutes.
• Tap-Tap-Tab: An autocomplete feature using Gemini Flash to suggest the next steps in a prompt, helping users articulate complex ideas.
• Agentic Engineering: A shift where developers partner with AI to build production-grade code, with senior engineers acting as stewards to ensure reliability and CI/CD compliance.

6. Notable Quotes

• Addy Osmani: "The history of software engineering is a history of a rising set of abstractions."
• Logan Kilpatrick: "We’re putting the means of opportunity in the hands of people who wouldn’t otherwise have been able to build this."
• Yasmin Ahmed: "We are seeing a rapid shift... from systems of intelligence to systems of action."

Synthesis/Conclusion

The conference underscored a fundamental shift in the AI landscape: the move from simple chatbot interfaces to autonomous, long-running agents capable of executing complex business processes. By integrating infrastructure (TPUs), data (Iceberg/Knowledge Catalog), and development tools (AI Studio/ADK), Google is lowering the barrier to entry for "vibe coding," enabling a new generation of builders to create sophisticated, production-ready software without needing traditional deep-coding expertise. The focus for the next 12 months is on reliability, governance, and inference efficiency to turn AI hype into measurable business ROI.