Cloudflare expands Agent Cloud

Cloudflare is widening its Agent Cloud push, packaging the pieces developers need to run artificial intelligence agents without stitching together their own backend. (blog.cloudflare.com) An AI agent is software that does more than answer one prompt: it keeps state, uses tools, waits for events, and resumes work later. Cloudflare’s Agents documentation says each agent runs on a Durable Object, a stateful micro-server with its own database, WebSocket connections, and scheduling. (developers.cloudflare.com) That setup is meant to replace the usual patchwork of stateless app servers, external databases, job queues, and session logic. Cloudflare says developers can deploy once on its network and scale to “tens of millions of instances,” while agents hibernate when idle and wake on demand. (developers.cloudflare.com) The company is tying the agent layer to other Cloudflare products it already sells. Agents can call models through Workers AI, which Cloudflare describes as serverless model inference on its network, and can hand off long-running, multi-step jobs to Workflows, which persists state for minutes, hours, days, or weeks with automatic retries. (developers.cloudflare.com; developers.cloudflare.com) Cloudflare is also building around Model Context Protocol, or MCP, a standard for connecting agents to tools and data sources. Its Agents guides now include hosted remote MCP patterns, and Cloudflare says its own managed MCP servers can expose more than 2,500 Cloudflare application programming interface endpoints through OAuth-protected access. (developers.cloudflare.com; developers.cloudflare.com) Security is part of the pitch because agents can take actions, not just generate text. Cloudflare One documentation says MCP portals can sit in front of multiple MCP servers, require login through Cloudflare Access, and log tool requests, with optional data loss prevention scanning through Gateway. (developers.cloudflare.com) The timing reflects a broader shift in cloud infrastructure marketing. Instead of selling only raw compute, storage, and networking, vendors are trying to offer managed runtimes for agent software that needs memory, permissions, orchestration, and audit trails in one place. (blog.cloudflare.com; developers.cloudflare.com) Cloudflare has been building toward this for more than a year through Workers, Durable Objects, Workers AI, and Workflows. Agent Cloud turns those separate services into a more direct pitch: build the agent as a class, connect it to models and tools, and let Cloudflare handle the machinery underneath. (developers.cloudflare.com; developers.cloudflare.com; developers.cloudflare.com)