Automating the Web With MCP: Infra That Doesn’t Break

Abstract

AI agents are only as strong as the infrastructure beneath them. In this talk, we’ll walk through the architecture behind Browserbase’s model context protocol (MCP), built to support stateful browser automation at scale. From concurrency models to session isolation, this talk dives into what it actually takes to run parallel agents in production, and what we learned the hard way. Expect live demos, real-world lessons, and infra blueprints you can steal.
 

Main Takeaways:

  • Infra-first thinking is crucial for agent-based web automation.
  • MCPs are not just tools, they’re runtime environments with real constraints.
  • Attendees will leave with:
    • Concrete patterns for scaling headless browser fleets.
    • Security tips for managing cookies and sessions safely.
    • Infra cost models and deployment blueprints.
    • Hard-won lessons from production-scale orchestration.

Interview:

What is the focus of your work these days?

We’re building a stateful, concurrency-safe browser infrastructure for AI agents. Our focus is making multi-agent orchestration reliable, secure, and scalable—without abstracting away the parts developers actually need to control, like context and memory.

What is the motivation behind your talk?

Everyone’s excited about agents but very few talks get into the gritty, technical realities of running browser-based agents at scale. We wanted to share the behind-the-scenes of how infrastructure decisions shape performance, reliability, and developer experience. If you’re building agents that touch the web, this is the stuff no one’s documenting yet—but should be.

Who is the target audience for this session?

Chief Innovation Officers, Engineering Leads, AI Engineers, Technical Leads, etc.


Speaker

Paul Klein

Founder @Browserbase, previously Director of Self-Service & Engineering Manager @Mux, Co-Founder & CTO @Stream Club, Technical Lead @Twilio Inc.

Paul Klein IV is a San‐Francisco‐based serial entrepreneur and engineer. After honing his chops at Twilio during its IPO and founding Stream Club—a live‐streaming platform acquired by Mux in 2021. In 2024 he launched Browserbase to give developers and AI agents fast, reliable, multi‐region headless‐browser infrastructure. In its first 16 months, Klein raised over $67 million (from investors like Notable Capital, Kleiner Perkins, CRV, and Okta Ventures). He views Browserbase as the “last‐mile” interface between large language models and the web, enabling end‐to‐end workflow automation far beyond traditional scraping.

Read more

From the same track

Session

Dynamic Moments: Weaving LLMs into Deep Personalization at DoorDash

Tuesday Nov 18 / 10:35AM PST

In this talk, we’ll walk through how DoorDash is redefining personalization by tightly integrating cutting-edge large language models (LLMs) with deep learning architectures such as Two-Tower Embeddings (TTE) and Multi-Task Multi-Label (MTML) models.

Speaker image - Sudeep Das

Sudeep Das

Head of Machine Learning and Artificial Intelligence, New Business Verticals @DoorDash, Previously Machine Learning Lead @Netflix, 15+ Years in Machine Learning

Speaker image - Pradeep Muthukrishnan

Pradeep Muthukrishnan

Head of Growth for New Business Verticals @DoorDash, Previously Founder & CEO @TrustedFor, 15+ Years in Machine Learning

Session

From Content to Agents: Scaling LLM Post-Training Through Real-World Applications and Simulation

Tuesday Nov 18 / 03:55PM PST

This talk presents a comprehensive journey through modern AI post-training techniques, from Pinterest's production-scale content discovery systems to enterprise agent training through Veris AI’s simulation.

Speaker image - Faye Zhang

Faye Zhang

Staff Software Engineer @Pinterest, Tech Lead on GenAI Search Traffic Projects, Speaker, Expert in AI/ML with a Strong Background in Large Distributed System

Speaker image - Andi Partovi

Andi Partovi

Co-Founder @Veris AI, Making AI Agents World-Ready

Session

Engineering at AI Speed: Lessons from the First Agentically Accelerated Software Project

Tuesday Nov 18 / 01:35PM PST

Claude Code is the first developer tool built specifically to maximize AI development velocity.

Speaker image - Adam Wolff

Adam Wolff

Engineer and Individual Contributor to Claude Code @Anthropic, Previously @Robinhood, @Facebook

Session

Deep Research for Enterprise: Unlocking Actionable Intelligence from Complex Enterprise Data with Agentic AI

Tuesday Nov 18 / 11:45AM PST

Deep Research as a consumer product redefined the AI space delivering true impact to many by searching through hundreds of websites, deeply thinking through the content, and generating a comprehensive report.

Speaker image - Vinaya Polamreddi

Vinaya Polamreddi

Staff ML Engineer; Agentic AI @Glean; Previously @Apple, @Meta, and @Stanford

Session

Improving Meta Generative Ad Text using Reinforcement Learning

Tuesday Nov 18 / 05:05PM PST

Reinforcement Learning with Performance Feedback (RLPF) unlocks a new way of turning generic GenAI models into customized models fine-tuned for specific tasks. This approach is especially powerful when combined with in-house data and performance metrics.

Speaker image - Alex Nikulkov

Alex Nikulkov

Research Scientist (RL lead for Monetization GenAI) @Meta