Web Agent Infrastructure Moves From Concept to Concrete Stack
The web is being re-architected for AI agents, and the shift is now visible in hard infrastructure rather than slideware. Google is pushing Chrome toward what it calls the “agentic web,” where browsers act as proactive assistants and websites expose structured capabilities instead of opaque pages. At the same time, Cloudflare has spent the past months assembling a six-layer platform for running agents, from compute to orchestration and browsing. Together, these initiatives are turning web agent infrastructure into a coherent stack: fast execution environments, agent-friendly browsing, and standardized ways for agents to discover and use site features. The goal is not just to let agents “browse like humans” but to let them transact, integrate, and automate against the web with predictable performance and semantics. This emerging baseline makes it much easier to talk seriously about AI agent performance and reliability at scale.
Cloudflare Browser Run Rebuild: Performance for Agent-Heavy Workloads
Cloudflare’s rebuilt Browser Run shows how quickly web agent traffic is reshaping infrastructure decisions. Originally sharing resources with human-focused Browser Isolation, Browser Run struggled with spiky, short-lived agent sessions that rapidly outpaced capacity. The team moved the service onto its dedicated Containers platform, added regional pools of pre-warmed browsers, switched state management from eventually consistent Workers KV to D1 with Queues, and collapsed multi-step WebSocket workflows into single HTTP calls executed entirely inside the container. The result: 4x higher concurrency, with support increasing from 30 to 120 simultaneous browsers per account, and 50% faster response times for quick actions. Support for WebGL and WebMCP further aligns Browser Run with modern web stacks. Importantly, existing users did not need to change their code, underlining that web agent infrastructure can be improved significantly beneath stable APIs.
WebMCP, DevTools Agents, and HTML-in-Canvas Bring Agents Into Chrome
Google is tackling web agent integration from the browser side with a set of features aimed at making web experiences first-class citizens in agent workflows. At the center is WebMCP, an emerging open standard that lets sites expose JavaScript functions, HTML forms, and other capabilities directly to agents instead of forcing them to scrape DOMs or interpret screenshots. Chrome is introducing WebMCP through an origin trial starting in version 149, giving developers a path to declare agent-facing tools in a structured, machine-usable way. Around this, Google is experimenting with DevTools agents and an HTML-in-Canvas API, which together make it easier to instrument, inspect, and render agent interactions with web content. Major brands like Booking.com, Expedia, Instacart, Intuit, Shopify, and Redfin are already exploring WebMCP, signalling that WebMCP Chrome support is likely to become a core part of web agent infrastructure in production environments.
Agent Readiness Score: Measuring the Web’s Legibility to Agents
Cloudflare’s Agent Readiness Score tries to answer a basic question: how legible is your website to AI agents? The public scanner at isitagentready.com fetches a site, runs 16 checks across five categories, and returns a composite 0–100 score with detailed pass/fail signals and AI-generated guidance. Checks span discoverability (robots.txt, sitemap, Link headers), content (Markdown negotiation for cheaper tokenization), bot access control (AI bot rules, Content Signals), and deeper API, auth, MCP, and skill discovery capabilities. The score is also exposed via Cloudflare Radar, an API, and even as a stateless MCP endpoint that agents themselves can call before deciding how to interact with a site. However, the composite agent readiness score can mislead if read in isolation: a content-only blog can score low because it lacks APIs or auth flows it never needed. The real value lies in aligning optimizations with your site’s purpose, not chasing a perfect number.
Toward Standardised, Performant Agent–Web Interactions
Taken together, Cloudflare’s platform work and Google’s Chrome roadmap signal a move toward standardised, performant agent–web interactions. On the infrastructure side, Cloudflare now offers a full-stack foundation for agents: Dynamic Workers for millisecond V8 execution, Sandboxes for full Linux containers with secure egress, high-concurrency browsing via Cloudflare Browser Run, and supporting primitives such as Agent Memory, Shared Dictionaries, and feature flags built for AI-generated code. On the browser and standards side, WebMCP Chrome support, DevTools agents, and HTML-in-Canvas aim to give agents a clear, sanctioned way to invoke site capabilities. The Agent Readiness Score bridges these layers by turning web “agentability” into something measurable, even if the optimization strategy matters more than the metric itself. As these pieces mature, building web agent infrastructure will feel less like bespoke engineering and more like assembling from interoperable, cross-vendor building blocks.