Cloudflare Agents Week: Sandboxes GA and Cloud 2.0

Cloudflare shipped more than a dozen products during its Agents Week developer event, led by Cloudflare Sandboxes reaching general availability as persistent, isolated Linux environments built on V8 isolates that launch in milliseconds with no container cold-start penalty. Cloudflare also raised Cloudflare Workflows limits from 4,500 to 50,000 concurrent instances, 100 to 300 creations per second, and 1 million to 2 million queued per workflow, and launched Managed OAuth with RFC 9728 compliance for agent authentication on behalf of users rather than shared service accounts. Cloudflare's AI Gateway now covers 70+ models across 12+ providers through a single Workers binding, while Code Mode in its MCP reference architecture cuts token usage by 81%.

Cloudflare just shipped more than a dozen products in a single week — and none of them are for building websites. Sandboxes hit general availability. Workflows scaled from 4,500 to 50,000 concurrent instances. Managed OAuth landed for non-human identities. The company called this “Cloud 2.0” and, for once, the marketing framing earns its keep. Agents Week was the company’s most consequential developer event since Cloudflare Workers launched in 2017. The name wasn’t accidental: this isn’t about adding AI features to an edge CDN. It’s a declaration that agents are a primary compute workload now, and that the infrastructure supporting them needs to be designed from scratch for that reality. Sandboxes Are GA — Agents Get a Real Computer The headline announcement is Cloudflare Sandboxes reaching general availability . A Sandbox is a persistent, isolated Linux environment: a real shell, a real filesystem, background processes, and state that survives across agent turns. An agent can clone a repository, install dependencies, run a test suite, and return to the same environment on the next interaction — without losing context. That sounds simple. It isn’t. Most “agent execution” today happens in ephemeral contexts: a function runs, returns output, and disappears. Building multi-step coding agents on ephemeral infrastructure is like asking someone to debug production code using a notepad and then burning it between each step. Sandboxes fix the fundamental memory and state problem for agent workloads. Sandboxes launch in milliseconds with no container cold-start penalty. They’re built on V8 isolates — the same foundation as Workers. You can try one now with cf sandbox create . Workflows: 11x the Scale The Cloudflare Workflows concurrency limits got a substantial overhaul. Previous ceiling: 4,500 concurrent instances, 100 creations per second, 1 million queued per workflow. New ceiling: 50,000 concurrent instances, 300 creations per second, 2 million queued per workflow . The control plane was rearchitected specifically for “durable background agents” — the kind that run for minutes or hours, not milliseconds. These numbers matter because agent workloads scale differently than web apps. A product with 100,000 users might trigger 100,000 simultaneous agents, each needing durable execution across multiple steps. The old limits weren’t a ceiling for edge cases; they were a ceiling for anyone serious about deploying agents at scale. Agent Identity: The Problem Everyone Is Ignoring The most underreported announcement from Agents Week is Managed OAuth with RFC 9728 compliance . Agents need to authenticate to internal systems — databases, APIs, SaaS tools — on behalf of users. The current industry answer is service accounts: create a credential, embed it in the agent, and hope nobody audits the blast radius when it leaks. Cloudflare’s approach follows RFC 9728 https://datatracker.ietf.org/doc/html/rfc9728 , the OAuth 2.0 Protected Resource Metadata standard. Agents authenticate on behalf of actual users, not synthetic service accounts. Internal apps become agent-ready with one configuration change. Non-human identity features — scannable API tokens, resource-scoped permissions — are now generally available as well. This matters most for enterprises. Every organization running agents against internal tools needs a governance story. “We created a shared service account” is not a governance story. AI Gateway: 70+ Models, 81% Fewer Tokens Cloudflare’s unified AI inference layer now covers 70+ models across 12+ providers — OpenAI, Anthropic, Google, Kimi K2.5, and others — through a single Workers binding. Switching between providers is a one-line code change. Cost monitoring is unified across all providers in one dashboard. The sharper story is Code Mode , a feature in Cloudflare’s MCP reference architecture that cuts token usage by 81%. Instead of agents issuing individual tool calls each with its own context overhead , Code Mode lets agents write TypeScript API calls directly. The cost savings on high-volume agentic deployments are substantial. As a concrete benchmark: using Kimi K2.5 through the AI Gateway for security review agents processing billions of tokens daily cut costs 77% versus mid-tier proprietary models. Is “Cloud 2.0” the Right Frame? Cloudflare CEO Matthew Prince put it plainly: “We are entering a world where agents are the ones writing and executing code.” The “Cloud 2.0” label is marketing, but the underlying argument is sound. Traditional cloud infrastructure was designed for servers that field requests from human users. Agent workloads are the inverse: many agents, acting in parallel, running long-duration tasks with state, identity, and compute requirements that don’t fit the request/response model. What Cloudflare shipped this week is infrastructure where agents are first-class citizens: they have homes Sandboxes , identities Managed OAuth , task persistence Workflows , memory Agent Memory , and safe compute environments Outbound Workers with zero-trust egress . That full stack didn’t exist twelve months ago. What to Try Now If you’re building AI agents, start with Sandboxes GA plus the rearchitected Workflows. Those two together give you persistent agent execution with durable multi-step orchestration. The Cloudflare Agents SDK https://developers.cloudflare.com/agents/ Project Think, currently in preview provides the application-layer framework on top. The full Agents Week release log is at the Cloudflare blog https://blog.cloudflare.com/agents-week-in-review/ . The Agents Week updates page https://www.cloudflare.com/agents-week/updates/ has every announcement organized by category — compute, security, tooling, and the agentic web.