Cloudflare expands Agent Cloud

Cloudflare used its Agents Week kickoff on April 12 to widen its Agent Cloud push, pitching its network as a place to run artificial intelligence agents that keep state, call tools, and stay connected in real time. (blog.cloudflare.com) An artificial intelligence agent is software that does more than answer one prompt: it remembers prior steps, uses external tools, and keeps working across sessions. Cloudflare’s Agents documentation says its TypeScript-based Agents Software Development Kit bundles memory, scheduling, tool use, coordination with other agents, and real-time connections into one class. (developers.cloudflare.com) The new tooling centers on Sandboxes, which Cloudflare described on April 13 as persistent, isolated environments with a shell, a filesystem, and background processes that start on demand and resume where they left off. The same post says developers can now use PTY terminal support, persistent code interpreters for Python, JavaScript, and TypeScript, filesystem watching, snapshots, and live preview URLs. (blog.cloudflare.com) Cloudflare is also exposing more of the control plane around those agents in its public docs. Its routing guide explains how requests are mapped to named agents, and its client software development kit docs say apps can connect over WebSockets or Hypertext Transfer Protocol for state sync, remote procedure calls, and streaming responses. (developers.cloudflare.com 1) (developers.cloudflare.com 2) Security is a central part of the pitch because agent software often runs model-generated code or handles user data. Cloudflare’s sandboxing post from late March said Dynamic Workers execute artificial intelligence-generated code in secure isolates with millisecond startup times, while its sandbox authentication post says credentials can be injected without giving the agent direct access to the secret itself. (blog.cloudflare.com 1) (blog.cloudflare.com 2) The company is tying that runtime story to its authorization layer. Cloudflare’s Model Context Protocol docs say its authorization flow uses a subset of OAuth 2.1 so users can log in and grant an agent or client access to account resources. (developers.cloudflare.com) This expands a strategy Cloudflare has been building in pieces for more than a year. In March 2025, the company introduced a JavaScript agents framework and Workers Artificial Intelligence updates, and in July 2025 it described OpenAI’s Agents Software Development Kit as the reasoning layer while Cloudflare supplied the runtime, identity, and state. (blog.cloudflare.com 1) (blog.cloudflare.com 2) Cloudflare’s recent releases show it wants developers to keep the full agent stack on one platform. On March 19, it said Workers Artificial Intelligence now runs Moonshot Artificial Intelligence’s Kimi K2.5 with a 256,000-token context window and multi-turn tool calling, and its workflows docs position Workflows as the background engine for long-running jobs while agents handle live interaction. (blog.cloudflare.com) (developers.cloudflare.com) The result is a more concrete version of Agent Cloud than last year’s marketing shorthand: not just model hosting, but named agents, persistent sessions, shell access, debugging surfaces, and permission controls published in product docs. Cloudflare is spending this week turning that package into a developer platform story with daily releases under the Agents Week banner. (blog.cloudflare.com)