Particle.news

Cloudflare Launches Agentic Cloud to Run Autonomous AI at Scale

The move bets that long‑running AI agents need purpose‑built runtimes, versioned storage, plus strict default security.

Overview

  • Cloudflare introduced its agentic cloud at Agents Week 2026, positioning the platform to host many long‑lived AI agents that work concurrently.
  • The company added new runtimes that span full operating‑system environments to fast isolates, plus Artifacts for Git‑style versioned storage and Sandboxes for persistent workspaces.
  • Security features now include Cloudflare Mesh for private network access, Managed OAuth that follows RFC 9728, scannable API tokens, and resource‑scoped permissions for least‑privilege access.
  • A unified inference layer connects to more than 14 model providers and the agent toolkit now covers inference, search, memory, a browser, a public‑beta email service, and an experimental real‑time voice pipeline, with Unweight compression to cut model size by up to 22%.
  • To help teams reach production, Cloudflare reworked Workflows for up to 50,000 concurrent runs and added a unified CLI, Local Explorer, feature flags, database integrations, a Registrar API, and a preview Rust‑based network stack called FL2.