Cloudflare ships AI agents

Cloudflare said on April 13 that it is expanding “Agent Cloud,” a set of tools for building, deploying, and scaling artificial intelligence agents on its network. (finance.yahoo.com) An agent is software that can take multi-step actions on its own, like calling tools, scheduling work, or updating systems instead of only answering a chat prompt. Cloudflare’s own documentation says its Agents software development kit is built for programs that remember state, stay connected in real time, and keep running across requests. (developers.cloudflare.com) Cloudflare said each agent runs on a Durable Object, its stateful micro-server product, with its own database, WebSocket connections, and scheduling. The company says that setup lets developers deploy once and scale to “tens of millions of instances” across Cloudflare’s global network. (developers.cloudflare.com) The new push is aimed at a basic bottleneck: agents often need a safe place to run code they generate on the fly. Cloudflare’s Dynamic Workers feature, announced in open beta on March 24, runs that code inside isolated sandboxes that the company says start in milliseconds and are faster and lighter than containers. (blog.cloudflare.com) Cloudflare framed the release as a move from laptop demos to production systems. In its April 13 announcement, the company said the package covers infrastructure, compute, deployment, and security for “millions of autonomous, long-running agents.” (finance.yahoo.com) That pitch builds on a strategy Cloudflare laid out on February 25, when it introduced its Agents software development kit and argued that agents need memory, tool use, and long-running execution rather than one-off chat responses. The company paired that earlier release with Workers Artificial Intelligence updates including structured outputs, tool calling, and longer context windows. (blog.cloudflare.com) Cloudflare is also trying to make its platform a distribution point for model providers, not just a hosting layer. Its Agents documentation says developers can use Workers Artificial Intelligence by default or swap in models from OpenAI, Anthropic, Google Gemini, or any service with an OpenAI-compatible interface. (developers.cloudflare.com) OpenAI joined the rollout on April 13, saying enterprises can access frontier models including GPT-5.4 inside Cloudflare Agent Cloud and deploy agents built on Codex harness to Cloudflare. OpenAI also said Codex harness is now generally available in Cloudflare Sandboxes. (openai.com) Cloudflare’s case against the older approach is mostly about cost and startup time. In its March 24 post, the company said Linux containers can take hundreds of milliseconds to boot, use hundreds of megabytes of memory, and often need to stay warm, which raises costs for agent workloads that spin up and down constantly. (blog.cloudflare.com) The company is betting that agents will look more like always-on services than chat windows. If developers buy that model, Cloudflare wants the place those agents live, remember, and act to be its network. (finance.yahoo.com)