> Markdown version of [/videos/1896-how-web-ai-can-power-the-agentic-web-jason-mayes-google?t=418](https://www.wearedevelopers.com/videos/1896-how-web-ai-can-power-the-agentic-web-jason-mayes-google?t=418). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md).

---

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

Jason Mayes bets his career that the browser is the future of agentic AI. Discover how local models give JavaScript engineers the same power as Python developers.

- **Speakers:** Jason Mayes
- **Event:** Coffee With Developers
- **Published:** May 27, 2026
- **Duration:** 20:57
- **URL:** https://www.wearedevelopers.com/videos/1896-how-web-ai-can-power-the-agentic-web-jason-mayes-google

## Summary

As artificial intelligence shifts toward an "agentic web," running AI models locally within the browser offers immense advantages for latency, cost reduction, and user privacy. Google's Web AI Lead, Jason Mayes, illustrates how developers can bridge the gap between heavy cloud operations and client-side processing by natively running smaller language models directly on consumer devices. Leveraging modern unified memory architectures, localized models open up powerful capabilities previously restricted by cloud computing costs or strict data compliance rules in sectors like medical and legal. Moving artificial intelligence into the frontend requires deliberate architectural choices between built-in browser APIs and the flexibility of deploying custom models via WebGPU. Today's hybrid implementations favor dynamic loading, providing a lightweight base model and dynamically injecting smaller Low-Rank Adaptation (LoRA) weights to alter agent capabilities without forcing extreme user downloads. Runtimes including TensorFlow.js, Microsoft's ONNX Runtime Web, and Google's RT.js seamlessly translate complex machine learning kernels into performant frontend experiences, moving past older WebGL limits and accelerating toward incoming WebNN standards. A pivotal breakthrough in making localized autonomous execution practical is the Model Context Protocol (WebMCP). Instead of burning expensive context windows by forcing algorithms to parse complex DOM trees, WebMCP allows website owners to expose specific, machine-readable data subsets securely. This structure authorizes agents to traverse logged-in sessions and execute multistep tasks automatically, like triaging hiring candidates directly through third-party platforms. Betting his career that the browser becomes a center point for agentic workflows, Mayes emphasizes that JavaScript engineers now hold as much power as Python backend developers to define the next evolution of intelligent web tooling.

**Keywords:** client-side web ai, agentic web browsers, hybrid ai architecture, webgpu model acceleration, webmcp implementation, lora weights tuning, built-in browser ai, dom token optimization, onnx runtime web, tensorflow.js performance, webnn api integration, local llm privacy, javascript machine learning, cross-domain model sharing

## Chapters

1. **The shift toward client-side agentic web development** (00:00) — Running small artificial intelligence models entirely within the web browser establishes the foundation for the future agentic internet.
1. **Adopting hybrid approaches for browser-based AI models** (00:53) — Testing local device capability determines whether an application should process data locally or leverage cloud fallbacks to optimize latency.
1. **Democratizing model execution via shared hardware architectures** (00:53) — Modern shared memory architectures allow standard laptops to allocate system RAM directly to graphics processors for high-speed local inference.
1. **Managing the user experience of large model downloads** (02:59) — Developers must balance the friction of heavy browser payloads against the complete privacy and zero-latency benefits of bundled local weights.
1. **Applying low-rank adaptation for efficient model updates** (04:26) — Pushing localized parameter adjustments over base models allows multiple personalities to exist concurrently without forcing massive sequential transfers.
1. **Distributing compute power through community GPU sharing** (05:18) — Leveraging peer-to-peer browser synchronization creates ad-hoc clusters that securely combine shared graphical horsepower for distributed querying workloads.
1. **Shaping open standards for web-based machine learning** (06:58) — The W3C actively solidifies standard capabilities like WebNN and WebMCP to guarantee future execution stability.
1. **Using the web browser as a central agentic hub** (07:52) — Supplying direct domain tool access to models positions the browser as the focal environment for orchestrating automated cross-application processes.
1. **Optimizing token usage with opt-in API exposure** (09:42) — Exposing limited subsets of programmatic endpoints actively restricts model access scopes while significantly reducing processing token expenditure.
1. **Addressing storage limits and model caching solutions** (11:36) — Fragmenting parameter files dynamically addresses native multi-gigabyte storage boundaries encountered internally by standard memory cache mechanisms.
1. **Navigating privacy concerns with proprietary browser AI** (13:30) — Employing open-source Apache licensed alternatives directly mitigates overarching security telemetry concerns associated natively with proprietary built-in models.
1. **Accelerating local inference using WebGPU and WebNN** (14:50) — Transitioning computational workloads from WebGL pipelines to WebGPU delivers significantly higher data throughput for dedicated mathematical calculation tasks.
1. **Choosing between built-in browser models and custom runtimes** (16:03) — Technical teams must reconcile the immediate speed of pre-installed web routines against the profound flexibility afforded by completely custom inference ports.
1. **Comparing ONNX runtime web and TensorFlow deployments** (18:02) — Integrating Python-based academia structures translates largely into adopting ONNX translation layers to bridge differences across historical JavaScript engines.
1. **Bringing JavaScript developers into web artificial intelligence** (19:00) — Utilizing locally verified toolchains allows frontend engineers to integrate sensitive enterprise systems without scaling traditional cloud backend infrastructure.

## Related Moments

- [The case for native AI in web browsers](https://www.wearedevelopers.com/videos/1572-privacy-first-in-browser-generative-ai-web-apps-offline-ready-future-proof-standards-based) (from "Privacy-first in-browser Generative AI web apps: offline-ready, future-proof, standards-based")
- [Leveraging Chrome AI and nano models for web applications](https://www.wearedevelopers.com/videos/1770-generate-ai-in-the-browser-with-chrome-ai-raymond-camden) (from "Generate AI in the Browser with Chrome AI - Raymond Camden")
- [Accelerating local machine learning models via WebNN](https://www.wearedevelopers.com/videos/100014-what-s-new-in-web-2026-edition) (from "What’s New in Web? 2026 Edition")
- [Exploring agentic browsers and artificial intelligence generation](https://www.wearedevelopers.com/videos/1723-wearedevelopers-live-graalvm-in-action-static-analysis-insights-and-more) (from "WeAreDevelopers LIVE - GraalVM in action, Static Analysis insights and more")
- [Defining web AI and spurring ecosystem innovation](https://www.wearedevelopers.com/videos/1296-exploring-the-future-of-web-ai-with-google) (from "Exploring the Future of Web AI with Google")
- [Real-world implementations of browser-based machine learning](https://www.wearedevelopers.com/videos/155-machine-learning-in-the-browser-with-tensorflowjs) (from "Machine learning in the browser with TensorFlowjs")

## Related Articles

- [The Web We Broke (And Why AI Agents Are Paying the Price) - AgentCon Berlin](https://www.wearedevelopers.com/magazine/735-the-web-we-broke-and-why-ai-agents-are-paying-the-price-agentcon-berlin)
- [MLOps And AI Driven Development](https://www.wearedevelopers.com/magazine/82-mlops-and-ai-driven-development)
- [WebMCP: Empowering Agents as First-Class Citizens of the Web](https://www.wearedevelopers.com/magazine/696-webmcp-empowering-agents-as-first-class-citizens-of-the-web)
- [Stephan Gillich - Bringing AI Everywhere](https://www.wearedevelopers.com/magazine/489-stephan-gillich-bringing-ai-everywhere)

## Related Jobs

- [AI Software Engineer (Germany)](https://www.wearedevelopers.com/jobs/48317-ai-software-engineer-germany) at **Sunhat**
- [AI Full Stack Engineer](https://www.wearedevelopers.com/jobs/ext/1354435-ai-full-stack-engineer) at **Almedia**
- [Principal Software Engineer, Enterprise AI Platform](https://www.wearedevelopers.com/jobs/ext/1467292-principal-software-engineer-enterprise-ai-platform) at **GitHub**
- [Staff Software Engineer, Copilot Experiences](https://www.wearedevelopers.com/jobs/ext/164361-staff-software-engineer-copilot-experiences) at **GitHub**
- [Head of AI Applications](https://www.wearedevelopers.com/jobs/ext/1231536-head-of-ai-applications) at **ZEISS Group**
- [Data Scientist](https://www.wearedevelopers.com/jobs/ext/1351648-data-scientist) at **Almedia**