Agentic Cloud: The Four Layers of Cloudflare's AI Strategy

On February 11, 2026, Cloudflare, the network and edge-computing company, reported its fourth-quarter 2025 results. Quarterly revenue was $614.5 million, up 34% year over year, and full-year 2026 guidance was $2.785 billion to $2.795 billion. In management's account AI was the main driver: Workers AI inference requests grew 4,000% year over year and AI Gateway requests grew 1,200%, and enterprise customers already supplied 73% of revenue. Two months later, from April 13 to 17, Cloudflare held Agents Week 2026, the densest product launch week in its history, on a single theme: redesign infrastructure around AI agents as a new kind of cloud workload. Taken together, what Cloudflare is doing breaks into four layers, from the most concrete product launches up to the most abstract strategic ambition. None of the layers stands alone, and each lays the foundation for the one above. Layer one, product: five rungs of execution for agent workloads The products from Agents Week 2026 are no scattered feature updates. Together they answer one question systematically: how should an agent run? Project Think is the next generation of the Agents SDK, upgrading agents from stateless orchestration to durable actor infrastructure. It has five core primitives: fibers (durable execution that can be interrupted and resumed), sub-agents, persistent sessions (state across calls), sandboxed code execution (for code the agent writes), and self-authored extensions (the agent writes tools to extend itself). Dynamic Workers answers a different question: where does the code an agent writes run? It is a fresh V8 isolate that cold-starts in milliseconds, wraps AI-generated JavaScript in a few megabytes of memory, and is destroyed when the run ends. Cloudflare's figures are 100 times the speed and 100 times the memory efficiency of containers. Each Dynamic Worker starts with almost no ambient permissions, and developers hand it capabilities one at a time through bindings, such as reading R2, reading a database, or calling an external API. Cloudflare Mesh covers how agents reach an enterprise's internal network. It gives enterprises a private network layer across clouds, so an agent on Workers can get fine-grained access to internal databases and private APIs through a VPC binding. An example of the granularity: a coding agent can read the staging database but cannot touch the production finance database. AI Search, formerly AutoRAG, packages the RAG pipeline into a plug-and-play primitive. Developers no longer stitch together chunking, embedding, vector indexing, keyword retrieval, and hybrid ranking. They upload files and AI Search runs the pipeline. It has R2 storage and a Vectorize index built in and connects to external models such as OpenAI and Anthropic through AI Gateway. Agent Lee is Cloudflare's own agent embedded in its dashboard, which understands the Workers, DNS, SSL, R2, and Cache resources in your account. It currently serves 18,000 users a day with 250,000 tool calls. It is both a product and a showcase of best practice, inside and outside the company. Lined up, these products fit a design Cloudflare calls the execution ladder. Tier 0 is Workspace (context with no sandbox), Tier 1 is Dynamic Worker (JS execution in a sandbox), Tier 2 adds npm integration, Tier 3 is a headless browser (web automation), and Tier 4 is a full OS sandbox with git, bash, and multi-language builds. Each action an agent takes is routed to the tier its security and capability needs call for. That tiering means Cloudflare no longer treats getting code to run as one problem. It treats it as a problem with five gradients, which marks a real effort to design for agent workloads as a new type of cloud load. Layer two, platform: CDN to Workers to AI Gateway to agent cloud, compounding