How Web AI Can Power the Agentic Web - Jason Mayes (Google)

Imagine an AI that orchestrates tasks across multiple sites, right from your browser. Learn how on-device AI is building a more powerful and private agentic web.

#1about 3 minutes

Running agentic AI models directly in the browser

A hybrid approach combining client-side models for privacy and latency with cloud fallbacks addresses current hardware limitations.

#2about 2 minutes

Solving the UX challenge of large model downloads

Using built-in browser AI or fine-tuning with smaller LoRA weights can avoid forcing users to download gigabytes of data.

#3about 2 minutes

Community innovations and W3C standardization efforts

Community projects are exploring peer-to-peer GPU sharing while the W3C works on standards like WebNL and WebMCP for a stable foundation.

#4about 4 minutes

How the browser can become a central agentic hub

The WebMCP standard allows websites to expose their functionality as tools, enabling agents to perform complex tasks across multiple services.

#5about 4 minutes

The technical foundations enabling modern web AI

Technologies like WebGPU and the upcoming WebNN provide the necessary hardware acceleration for running complex models efficiently in the browser.

#6about 3 minutes

Comparing web AI runtimes and implementation approaches

Developers can choose between built-in browser AI or more flexible "bring your own model" approaches using runtimes like TensorFlow.js and ONNX.

#7about 2 minutes

A call to action for JavaScript developers

JavaScript developers are encouraged to explore Web AI to build privacy-focused applications and reduce long-term costs by shifting computation to the client.