Google I/O 2026 · Keynote Summary

The Agentic Web is Here.

From passive chat interfaces to proactive autonomous systems. Developer guidelines and platform highlights.

The Paradigm Shift

Moving from chat windows to active agency.

Web products are transitioning from conversational bots to goal-oriented background runtimes. Tomorrow's software navigates, coordinates, and transacts directly on behalf of users.

10M

token context window in Gemini 3.0 Ultra, enabling the live ingestion of entire codebases, database tables, and hours of video.

Technical Architecture

Three pillars of the agentic web platform.

Local Agent Execution

Browsers run low-latency planning loops natively. Local runtime constraints mean user data never leaves the browser scope.

Semantic Web Standards

A standard schema representation for buttons, inputs, and tables replaces fragile scraper scripts with structured action protocols.

Secure Credentials

Encrypted auth tokens enable agents to checkout and complete financial transactions securely without exposing raw secrets.

Execution Flow

Four stages of executing an autonomous web task.

01

Goal Intent

The user specifies a high-level outcome in natural language (e.g. "Reconcile Q2 travel expense reports").

02

Plan Graph

The local agent parses the target site schemas and creates a deterministic navigation path.

03

Navigate

The agent utilizes a Chrome-native virtual viewport to input data and submit forms without UI latency.

04

Verification

A summary card is rendered to the user to confirm costs, dates, and destinations before checkout.

10x

reduction in planning latency by running the initial task-graph evaluation on-device via Google Chrome's native Gemini Nano.

Design Evolution

From static funnels to dynamic agent-friendly UI.

Before · Human-Only UIs

Multistep forms and complex nested navigation.

Websites designed exclusively for manual cursor clicks. High cognitive load, long session times, and high user-drop rates during checkout.

After · Hybrid Interfaces

Structured semantic layers built for agent interactions.

Websites expose lightweight JSON schemas. The UI dynamically collapses to display only high-priority user verification screens.

Mobile Ecosystem

Android 17: Platform-level agency.

App Actions API

System-wide agents deep-link directly into third-party apps, skipping screens to complete workflows instantly.

Private Compute Core

Sensitive context profiles remain sandboxed. Permissions are managed at the operating-system level.

Multimodal Reading

The on-device model reads screen context dynamically, providing real-time accessibility and task automation.

"

The browser is no longer just a page viewer — it is a secure execution environment for autonomous agents.

— Sundar Pichai, CEO at Alphabet & Google

Browser Platforms

Chrome becomes a local machine learning runtime.

By leveraging WebGPU and WebNN, developers can run models up to 7B parameters at native execution speeds on client GPUs. This model bypasses expensive server bottlenecks and guarantees immediate offline functionality.

4.8EFL

of dedicated machine learning compute in TPU v6 clusters, scaling developer model training velocities by 3.2x over TPU v5.

Tooling Suite

Ready-to-use developer platforms.

Firebase AI Logic

Deploy serverless logic that securely runs and orchestrates agent planning graphs in clean, isolated environments.

Material 4 Systems

A revised component framework featuring built-in state nodes for agentic queries and screen readers.

Project IDX IDE

A cloud developer workspace with automated integration testing for agent-driven user flows and APIs.

"

Material 4 is not about colors or rounded corners. It is an adaptive canvas designed for humans and agents to co-create UI in real time.

— Google User Experience Design Lead

Get Started

The agentic web is live today.

Access the Gemini 3.0 SDK, Material 4 spec, and Android 17 SDK in the developer portal at developer.google.com/io/2026.

1 / 14
← / → · scroll · swipe