Cloudflare (NYSE: NET) capped its Agents Week on April 15 with the densest product launch sequence the company has ever aimed at a single audience: AI agent developers. Four new products shipped in a single afternoon, extending the Sandboxes GA and Dynamic Workers announced earlier this week.
The headline additions: an Agents SDK v2 preview described as “batteries-included” for agents that “think, act, and persist”; Agent Lee, an in-dashboard agent that replaces manual tab navigation with natural language commands; Cloudflare Mesh, a private networking layer for agents accessing internal infrastructure; and cf, a unified CLI spanning all 3,000 Cloudflare API operations, explicitly co-designed for both human developers and AI agents.
“The way people build software is fundamentally changing. We are entering a world where agents are the ones writing and executing code,” CEO Matthew Prince told SiliconANGLE. “Today, we are making Cloudflare the definitive platform for the agentic web.”
The Stack, Assembled
Agents Week ran April 13 through 15. The earlier days delivered Sandboxes GA (persistent, isolated Linux environments for agents), Dynamic Workers (isolate-based compute spinning up in milliseconds), Artifacts (Git-compatible storage), and Think (a persistence framework for multi-step tasks). Tuesday’s additions complete the picture.
The Agents SDK v2 preview moves Cloudflare’s agent toolkit from lightweight primitives to a full runtime. According to the Agents Week kickoff post, the SDK targets “long-running, multi-step tasks rather than just responding to single prompts,” positioning it as direct competition for OpenAI’s Agents SDK and LangChain’s orchestration layer.
Agent Lee is Cloudflare eating its own product. The company deployed an agent inside its own dashboard that uses sandboxed TypeScript to troubleshoot, diagnose, and manage stack configurations through natural language. Instead of navigating dozens of tabs to check DNS records, firewall rules, and Worker configurations, users describe what they need.
Mesh Brings MCP Governance
The most architecturally significant release may be Cloudflare Mesh. The product extends Cloudflare’s existing Zero Trust platform (WARP Connector, Gateway policies, Access rules) to cover autonomous agent traffic. Agents built with the Agents SDK can reach private databases, internal APIs, and staging environments through the same policy framework that governs human access.
For teams running MCP servers internally, Mesh addresses a gap that has been widening for months: MCP’s protocol lacks native authorization controls. Cloudflare’s answer is to layer its existing Access, AI Gateway, and MCP server portals on top of Mesh, creating what amounts to the first opinionated MCP governance architecture from a major infrastructure vendor.
“Agents need to reach private resources, but the tools for doing that were built for humans, not autonomous software,” the Mesh blog post states. “VPNs require interactive login. SSH tunnels require manual setup. Exposing services publicly is a security risk.”
The Scale Problem
Cloudflare’s framing of Agents Week rests on a specific math problem. The kickoff post lays it out: if 100 million U.S. knowledge workers each ran an agentic assistant at 15% concurrency, that requires capacity for 24 million simultaneous sessions. At 25 to 50 users per CPU, that demands 500,000 to 1 million server CPUs for a single country with a single agent per person.
Cloudflare’s argument is that containers cannot serve this scale economically. V8 isolates, the foundation of Cloudflare Workers for eight years, spin up in milliseconds and run at what Cloudflare claims is 100x the speed and a fraction of the cost of container-based alternatives. The Agents Week products layer persistent compute (Sandboxes), networking (Mesh), developer tooling (cf CLI), and an SDK on top of that isolate foundation.
Figma is already running agents on the platform. Kate Reznykova, writing in the Sandboxes GA announcement, quoted Figma’s team describing their use of Cloudflare containers for Figma Make, their agent-powered design tool.
The Convergence With OpenAI
The timing is notable. On the same afternoon Cloudflare shipped Sandboxes GA and its SDK v2 preview, OpenAI independently released sandboxing capabilities for its own Agents SDK. Two of the most important agent infrastructure platforms in the ecosystem arrived at the same architectural conclusion on the same day: unsandboxed agents in production are a liability. Isolated execution environments are now the baseline.
OpenAI’s SDK update explicitly names Cloudflare as one of its supported sandbox providers, alongside E2B, Modal, Daytona, and others. The two platforms are complementary, not competing: OpenAI provides the model and orchestration layer, Cloudflare provides the compute substrate.
What Builders Get Now
The full Agents Week stack is available through a standard Cloudflare Workers developer account. Sandboxes are GA. Dynamic Workers are in open beta. The Agents SDK v2 is in preview. Mesh is available to existing Cloudflare One customers and new signups. The cf CLI is shipping today.
For teams already running Cloudflare Workers, the migration path is incremental: add Sandboxes when agents need persistent compute, add Mesh when agents need private network access, upgrade to SDK v2 when the preview stabilizes. For teams evaluating agent infrastructure from scratch, Cloudflare just made the case that one vendor covers compute, networking, storage, security, and developer tooling in a single integrated platform.