Cloudflare ships AI agent sandboxes. Edge compute meets agent runtime.

Cloudflare ran Agents Week in late April, shipping a set of primitives aimed squarely at AI-agent workloads [Cloudflare blog].

── What shipped ──

Cloudflare Sandboxes. Persistent, isolated execution environments that give an agent a real computer — shell, filesystem, background processes. They start on demand, scale to millions, and "pick up exactly where they left off" between invocations. This is the missing piece between "stateless serverless functions" and "I need to run npm install before my agent can do anything."

Unified inference layer. A single API surface that fronts 14+ model providers — Anthropic, OpenAI, Google, Mistral, Meta, and others. Workers binding for running third-party models, plus an expanded multimodal catalogue [Agents Week updates].

Network expansion. Cloudflare's edge is now at 320 global points of presence, up from 200 in 2024. Workers runtime 2026.3 added long-lived TCP connections and 16GB per-node memory.

── Why it matters ──

The Sandboxes piece fills a real gap. Existing AI-agent platforms (Devin, Replit Agent, Lovable) bundle their own execution sandbox. Anyone building a custom agent has had to either run their own VM infrastructure or accept the limits of stateless functions. Sandboxes make "real computer" a one-line dependency.

The unified inference layer is more interesting strategically. Vercel has been pushing in this direction with their AI SDK, but as a client-side abstraction. Cloudflare doing it server-side, at the edge, with bring-your-own-key support, lets developers swap models without changing deployment code.

For shipping engineers, two practical effects:

• Self-hosted agent platforms become cheaper to build. Sandboxes + unified inference + Workers replaces what previously required a half-dozen vendors.
• Latency-sensitive agent loops get tighter. With model invocation and sandbox execution co-located on Cloudflare's edge, round-trip times for tool-using agents drop materially.

── Editor's take ──

Cloudflare's bet is that the agent runtime — not the model — is where the durable infrastructure value sits. Sandboxes are the most concrete expression of that bet so far. If you're building agents, this is now a credible alternative to running your own infrastructure on hyperscalers.