From Browser Isolation to a Purpose-Built Agent Infrastructure Stack
Cloudflare’s work on AI agents started with a simple observation: Browser Run, originally sharing infrastructure with Browser Isolation, was being pushed beyond its design by agent traffic. Human browsing sessions are long and steady; AI agent usage is short, spiky, and massively concurrent. That mismatch triggered a broader rethink that ended with a complete agent infrastructure stack, not just a single feature. Over the past two months, Cloudflare has delivered six infrastructure primitives that collectively cover compute, orchestration, state, networking, and security needs for AI agent deployment. The goal is to give developers a vertically integrated platform where agents can not only browse the web, but also reason, call tools, store memory, and safely interact with external systems. Instead of gluing together disparate services, teams can now assemble production-grade agents on one stack optimized for high-frequency, automated workloads.
Inside the Six-Layer Agent Infrastructure Stack
At the foundation of Cloudflare’s agent infrastructure stack is compute, split into two tiers. Dynamic Workers offer V8 isolate-based execution that starts in milliseconds, ideal for lightweight tasks such as linting, type checking, and fast API calls. For heavier jobs, Sandboxes provide full Linux containers so agents can run git, bash, development servers, or multi-language builds. Credentials are injected securely via an egress proxy so agents never see raw tokens. Above compute, Cloudflare layers orchestration, storage, networking, and security primitives tailored to AI agents. Dynamic Workflows coordinate multi-step tasks, while services like Agent Memory and Shared Dictionaries give agents persistent state across runs. Features such as Redirects for AI Training and Unweight focus on how content is exposed and compressed for models. The result is an agent infrastructure stack that addresses not just execution, but also data access, routing, and governance in one coherent platform.
Browser Run Rebuilt on Containers: More Concurrency, Lower Latency
Browser Run acts as the browsing layer of Cloudflare’s agent infrastructure stack, and its recent rebuild on the Containers platform was crucial for scale. Previously, Browser Run shared infrastructure with Browser Isolation, where long-lived human sessions coexisted awkwardly with bursty agent activity. The new design dedicates infrastructure to agents, using regional pools of pre-warmed browsers running in containers. Concurrency has jumped from 30 to 120 simultaneous browsers, a 4x improvement, while quick actions now respond about 50% faster. Under the hood, state management moved from Workers KV, which offered eventual consistency and could cause race conditions, to D1 plus Queues, enabling transactional assignment and batch writes for up to 500,000 containers per location. Cloudflare also removed multi-step WebSocket flows for quick actions, replacing them with single HTTP requests executed entirely within the container, simplifying integration and cutting latency for AI agent deployment.
Agent Readiness Score: Useful Signal, Dangerous North Star
Cloudflare’s Agent Readiness Score turns the fuzzy idea of “agent-friendly” websites into a concrete metric. Through isitagentready.com, Radar, an API, and an MCP endpoint, developers can scan any URL and receive a composite score plus a breakdown across 16 checks in five categories, including discoverability, content, bot access control, and API/Auth/MCP/skill discovery. The scanner looks for signals such as robots.txt and sitemaps, Markdown content negotiation, AI bot rules, Web Bot Auth, API catalogs, OAuth discovery, MCP server cards, and agent skills indexes. However, the composite agent readiness score can mislead if taken at face value. A content-only blog, for example, may score poorly in API or commerce-related checks it simply does not need, even if it is perfectly usable by agents. Optimizing purely for a higher score risks Goodhart’s law: overfitting to the metric instead of improving real agent experiences.
What This Stack Enables for AI Agent Deployment at Scale
When combined, Cloudflare’s six-layer agent infrastructure stack and the rebuilt Cloudflare Browser Run unlock a new level of autonomy for AI agents. Agents can programmatically scan a site’s agent readiness, negotiate content formats, and discover APIs, auth flows, and MCP endpoints. On the backend, they can execute code in Dynamic Workers and Sandboxes, persist memory, and orchestrate multi-step workflows. This makes it feasible for agents to handle complex lifecycle tasks end-to-end: creating cloud accounts, registering domains, deploying applications to production, and managing subscriptions and configuration over time. For developers, this means less time wiring together bespoke infrastructure and more time focusing on agent logic and safety constraints. The challenge is to treat tools like the Agent Readiness Score as guidance rather than targets, and to design agents that leverage this stack for robustness, observability, and secure, large-scale automation.
