Cloudflare and OpenAI Launch Agent Cloud, Giving Enterprises Browser‑Run AI Agents

Cloudflare’s new Browser Run service now runs full Chrome sessions on its edge network, letting AI agents act like real users, the company blog says. The feature adds a live‑view console so operators can watch an agent fill forms, click links or capture screenshots in real time, and intervene instantly if something goes wrong.

The platform also exposes the Chrome DevTools Protocol directly, giving agents granular control over navigation and page rendering, according to the same blog post. Developers can record every session for later debugging, and replay them to fine‑tune workflows without rebuilding the entire script.

Human‑in‑the‑loop support lets an agent hand off to a person when it hits a roadblock such as a login challenge or unexpected UI change. After the human resolves the issue, control returns to the AI, preventing outright failures, the blog notes.

Cloudflare says the service now supports MCP clients like Claude Desktop, Cursor and OpenCode, enabling those coding assistants to use Browser Run as a remote browser. WebMCP, another addition, lets websites publish machine‑readable action maps so agents can discover and invoke supported functions reliably.

For enterprises, the launch pairs Browser Run with OpenAI’s GPT‑5.4 model and isolates‑based Dynamic Workers, a combination that “challenges containers as the default runtime for enterprise AI agents,” Forbes reports. The integration promises higher security and lower latency for large‑scale, browser‑driven AI workloads.