From Capacity Crunch to 4x Concurrency in Browser Run
Cloudflare’s Browser Run was originally built on the same infrastructure as its Browser Isolation service, where long-lived human browsing sessions sit uneasily alongside the short, spiky workloads of AI agents. As agent builders discovered the product, demand quickly exceeded what the shared platform could sustain. In response, Cloudflare rebuilt Browser Run on top of its own Cloudflare containers platform, separating it from Browser Isolation and tuning it specifically for web agent infrastructure. The result: support for 120 simultaneous browser sessions per location, up from 30, delivering 4x higher concurrency with no changes required from existing users. Quick actions now respond around 50% faster, and the system adds support for WebGL and WebMCP, expanding the kinds of web apps agents can drive. For enterprises struggling with browser automation scaling, this new foundation is designed to handle production-level traffic rather than experimental pilots.
Inside the Six-Layer Cloudflare Agent Infrastructure Stack
The Browser Run rebuild is only the visible tip of a broader six-layer platform Cloudflare has assembled for web agents. At the compute tier, Dynamic Workers provide ultra-fast V8 isolate execution for light tasks like API calls, while Sandboxes supply full Linux containers when agents need tools such as git, bash, or dev servers. Both are wrapped in security features like credential-safe egress, ensuring agents never directly handle raw tokens. On top sits orchestration via Dynamic Workflows, an open, roughly 300-line engine intended to coordinate multi-step agent flows rather than lock teams into a heavyweight framework. Browser Run becomes the browsing layer in this stack, giving agents a high-concurrency, low-latency way to interact with websites. Together, these layers form a web agent infrastructure that aims to cover everything from code execution to workflow management and browser automation at scale.
Why Moving Browser Run to Dedicated Containers Matters
The architectural changes behind Browser Run’s performance jump are central to its enterprise relevance. By moving to dedicated regional pools of pre-warmed containers, Cloudflare can spin up browsers quickly without competing with interactive user sessions. State management has been migrated from an eventually consistent key-value store to a transactional combination of D1 and Queues, enabling reliable assignment and batch operations for up to 500,000 containers per location. For quick actions, Cloudflare removed the previous multi-step WebSocket choreography and now executes them as single HTTP requests fully inside the container. This not only reduces latency but also simplifies integration for automation platforms that prefer request–response patterns. Combined, these changes directly address long-standing production pain points in browser automation scaling: unpredictable cold starts, race conditions under load, and complex networking flows that are difficult to operate at enterprise volume.
Agent Readiness Score: Useful Signal, Dangerous Composite Metric
Alongside its runtime and container work, Cloudflare also introduced the Agent Readiness Score at isitagentready.com, a scanner that evaluates how prepared a website is for AI agents. It tests 16 signals across five categories, from basic discoverability (robots.txt, sitemaps, link headers) and content negotiation for Markdown, through bot access control and emerging standards such as Web Bot Auth and WebMCP. The score is helpful for quantifying whether a site is legible to modern web agent infrastructure and can be accessed via UI, API, or as an MCP tool that agents themselves call. Yet the composite number can mislead teams. A content-only site can score poorly because it correctly omits OAuth, API catalogs, or commerce flows it does not need. Different presets can double the score without any real change in content. For enterprises, the lesson is clear: treat the agent readiness score as diagnostic detail, not a vanity metric.
What This Stack Means for Production-Grade Browser Automation
Taken together, Cloudflare’s six-layer platform, rebuilt Browser Run, and Agent Readiness Score mark a shift from experimental agent demos to production-grade web automation. Enterprises have historically struggled to run browser-based automations reliably at scale, battling concurrency ceilings, inconsistent page access, and fragile orchestration logic. The new Cloudflare containers-based Browser Run aims to remove infrastructure headroom as the limiting factor, while Dynamic Workers and Sandboxes provide flexible execution environments tied into the same network edge. On the other side of the wire, the readiness scanner pushes website owners toward consistent, machine-readable interfaces that agents can parse and authenticate against. The combination offers a more complete web agent infrastructure: from how agents run and coordinate, to how they discover, access, and automate real-world applications. For teams planning large fleets of agents, the focus now shifts from raw capacity to designing workflows and governance on top of a more capable substrate.
