Cloudflare becomes the hands for Claude Managed Agents!

Hey everyone, it's Shiichan! Today I found a fun announcement about splitting an agent's "brain" from its "hands." Let me walk you through it!

Cloudflare Blog blog.cloudflare.com

What was announced?

On the Cloudflare Blog, Cloudflare and Anthropic announced an integration that connects Anthropic's Claude Managed Agents with Cloudflare Sandboxes. In short: you run the agent loop (the thinking part) on the Claude Platform, while Cloudflare handles code execution, securing connections, and custom tool calls.

Anthropic describes this as "decoupling the brain from the hands." The agent logic lives on Anthropic's platform, and the execution infrastructure lives on Cloudflare, so the two sides split the work.

Why it matters

Once you actually want an agent to do real work, it needs "hands": running code, reaching internal services, driving a browser, and so on. Building all of that securely and at scale yourself is hard. With this integration, Cloudflare takes care of that hands-side infrastructure, so developers can run lots of agents while keeping control over security, compliance, and performance.

What changes

There's a deployment template ready to go. You fork it, write your wrangler config, deploy, and you immediately get a bunch of capabilities:

• Outbound proxies that inject credentials so secrets are never exposed to the agent sandboxes (and data exfiltration is prevented)
• Detailed sandbox metrics, logs, SSH access, and custom images
• Private connectivity to internal services without exposing them to the Internet
• Browser control with session recording
• An individual email address for each agent

The control plane runs on Workers, and state is persisted with Durable Objects, so an agent can pick up where it left off even if a session is interrupted.

Dive Deep

There are two kinds of sandboxes, and you pick based on your needs:

• microVM type: a stateful Linux microVM with the full Linux tool ecosystem. Great when you want to do everything, like building applications.
• Isolate type (Dynamic Workers): built on V8 isolates, with millisecond boot times. It runs arbitrary code via Codemode and has a file system, scales to tens of thousands of concurrent agents, and is cheaper and faster. If you don't need full Linux, this is the way.

Security is thorough too. Connections to external services go through customizable outbound proxies, with zero-trust credential injection, post-quantum encrypted networking, per-tenant and per-agent egress policies, and service allow/deny lists.

There are plenty of built-in tools: browser_search, browser_execute, screenshot, and browse for the browser; send_email, email_read, and email_list for email; call_service to reach private services over Mesh or VPC; image_generate for images via Workers AI; and web_fetch that fetches through a Cloudflare-hosted browser.

You can add your own tools easily with zod schemas. Fork the repo, define a schema in custom-tools.js, bind resources in your wrangler config, and deploy — no extra infrastructure to run yourself.

Observability isn't left out either: the control plane has a UI to track sandbox state, logs can be shipped to external providers like Datadog or Splunk, and you can SSH into a running machine to poke at it interactively.

Wrap-up

• From the Cloudflare Blog: an integration where Cloudflare Sandboxes become the "hands" for Claude Managed Agents
• The brain (agent loop) runs on the Claude Platform, the execution infra on Cloudflare
• A Workers control plane plus Durable Objects persist state and recover from interruptions
• Choose between microVM sandboxes (full Linux) and Isolate sandboxes (millisecond boot, tens of thousands concurrent, cheap)
• Credential-injecting proxies and post-quantum encryption keep it secure, and you can add your own tools with custom-tools.js and zod

This one is for people who want to scale lots of agents securely and cheaply, connect to internal services without exposing them to the Internet, and keep a close eye on what their agents are doing!