From Browser Isolation to a Dedicated Cloudflare Agent Infrastructure
Browser Run started as an offshoot of Cloudflare’s Browser Isolation product, sharing infrastructure that was tuned for long-lived human sessions. As AI agent development accelerated, that design hit a wall: agents generate short, spiky bursts of traffic that rapidly exceeded capacity. Cloudflare’s response was to rebuild Browser Run on its own Containers platform, giving web agents their own optimized execution environment instead of competing with interactive users. This shift is more than a scaling tweak; it is the capstone of a broader Cloudflare agent infrastructure push. Over the past two months, Cloudflare has assembled a six-layer stack that spans compute, orchestration, data, browsing, and observability for agents. Browser Run now sits as the dedicated browsing layer in that stack, designed from the ground up for automated, high-volume web interactions rather than human-driven browsing sessions.
Browser Run Performance: Containers, Concurrency, and Faster Agent Feedback
Rebuilding Browser Run on Containers delivered measurable performance gains that matter directly for Browser Run performance in production agent systems. Concurrency has jumped from 30 to 120 simultaneous browsers, a 4x increase that lets more agents run parallel browsing tasks without queueing. Quick actions now return roughly 50% faster, which compounds across multi-step agent workflows where every second of latency affects perceived intelligence. Under the hood, Cloudflare moved from shared Browser Isolation infrastructure to regional pools of pre-warmed containers, so new sessions start fast. State management migrated from Workers KV—where eventual consistency could trigger race conditions—to a D1 plus Queues model that supports transactional assignment and batch writes for hundreds of thousands of containers per location. Cloudflare also removed the multi-step WebSocket choreography for quick actions, replacing it with single HTTP requests executed entirely inside the container, simplifying integration and reducing overhead for developers.
The Six-Layer Stack: A Full Platform for AI Agent Development
Cloudflare now positions Browser Run as one piece of a six-layer Cloudflare agent infrastructure stack aimed at end-to-end AI agent development. At the compute layer, Dynamic Workers offer lightweight V8 isolate execution that boots in milliseconds for linting, type checking, and API calls. For heavier workloads, Sandboxes provide full Linux containers, enabling tools like git, bash, dev servers, and polyglot build chains, with credentials injected securely via an egress proxy so agents never see raw tokens. Above compute, orchestration is handled by Dynamic Workflows, an open, MIT-licensed workflow engine tuned for agent-centric patterns. Data features such as Agent Memory, Shared Dictionaries, and Redirects for AI Training give agents durable context and control over training signals. Browser Run provides the browsing layer, while features like Unweight and Flagship round out observability, optimization, and safe rollout of AI-generated changes. Together, these layers turn Cloudflare’s network into a vertically integrated runtime for web-native agents.
Agent Readiness Score: A Useful but Easy-to-Misread Metric
Parallel to the runtime stack, Cloudflare introduced the Agent Readiness Score via isitagentready.com, a scanner that tests how legible a website is to agents. It inspects 16 checks, grouped into five categories: Discoverability, Content, Bot Access Control, API/Auth/MCP & Skill Discovery, and Commerce. The scanner can be invoked via web UI, via the URL Scanner API, inside Cloudflare Radar, or as a stateless MCP server that agents can call as a tool before deciding how to interact with a site. However, optimizing purely for the composite agent readiness score is risky. The score measures delivery infrastructure, not business intent or content quality, and different presets can yield very different numbers for the same site. A content-only blog, for example, may legitimately skip OAuth or MCP endpoints yet still score poorly overall. Developers should treat the score as a diagnostic surface to prioritize relevant fixes, not as a leaderboard metric to chase at all costs.
What This Means for Developers Building Web-Scale Agents
For developers, Cloudflare’s completed stack changes both how you build and how you scale AI agents. Browser Run’s improved concurrency and latency enable more ambitious browsing-heavy workflows—think research agents, monitoring bots, or autonomous QA systems—that can fan out across many pages without collapsing under session limits. The combination of Dynamic Workers, Sandboxes, and Dynamic Workflows offers a continuum from ultra-fast function-style tasks to long-running container workloads, orchestrated as a single agent system. On the web-facing side, the Agent Readiness Score and its underlying checks effectively define a practical checklist for making your properties agent-friendly: robots.txt hygiene, sitemaps, Markdown negotiation, AI bot rules, and emerging standards like API catalogs, MCP server cards, and agent skills indexes. The strategic move is to align your infrastructure with real agent use cases—access patterns, authentication, and capabilities—rather than gaming a single composite score, leveraging Cloudflare’s platform to keep those interactions both efficient and scalable.
