Cloudflare and GitHub launch sandboxes

Cloudflare and Anthropic moved on May 19 to let developers run Claude Managed Agents inside Cloudflare-hosted execution environments, and GitHub followed on June 2 by putting cloud and local sandboxes for Copilot into public preview. Both launches center on the same problem: how to let coding agents run commands, touch files and reach services without giving them broad, persistent access to a developer’s machine or company network. Cloudflare said its service gives Claude agents isolated execution, private-service connectivity and observability controls, while GitHub said its sandboxes provide a native layer for secure tool execution with identity, governance and policy controls built in. ### Why are Cloudflare, Anthropic and GitHub all talking about sandboxes now? Anthropic said last month that managed agents separate an agent into a session, a harness and a sandbox, with the sandbox serving as the execution environment where Claude can run code and edit files. That design lets the execution layer be swapped without changing the rest of the agent stack, according to Anthropic’s engineering post. GitHub said on June 2 that “secure execution environments become foundational infrastructure” as AI agents take on a larger role in software development. (blog.cloudflare.com) Cloudflare used similar language in its May 19 announcement, saying organizations can run core agent loops on Claude while using Cloudflare’s network and Workers platform to execute code, secure private connections and provide specialized tools. ### What exactly did Cloudflare launch for Claude Managed Agents? (anthropic.com) Cloudflare said its integration with Anthropic gives developers a “fast, isolated execution environment for autonomous code delivery” and lets builders scale agent workflows globally while controlling access to private backends. The company said developers can customize tools and runtimes, inspect sandbox metrics and logs, SSH into running machines, and choose between traditional microVMs and lightweight isolates for untrusted code. (github.blog) May 19 materials from Cloudflare said the product is called Cloudflare Environments for Claude Managed Agents. In a press release, Cloudflare said the setup allows organizations to execute code on Cloudflare’s global network, secure private connections and extend agents with specialized tools on the Workers developer platform. ### What did GitHub put into preview for Copilot? GitHub said on June 2 that cloud and local sandboxes for GitHub Copilot are now in public preview. (blog.cloudflare.com) The company’s documentation says the sandboxes are isolated environments that let Copilot interact with code, tools, filesystem and network resources either on a local machine or in fully isolated cloud environments. GitHub’s changelog said local sandboxing can be turned on inside a Copilot session with `/sandbox enable`. (cloudflare.com) The company said the feature is meant to give Copilot a secure execution layer with consistent identity, governance and policy controls across local and cloud use. ### What do these products have in common? Both companies are packaging execution as a bounded service instead of assuming an agent should run directly against a developer workstation or a flat corporate network. (github.blog) Cloudflare emphasized isolated per-agent environments, controlled connectivity to private services and native observability. GitHub emphasized isolated execution, policy controls and secure interaction with files, tools and network resources. The overlap suggests that orchestration around the sandbox is becoming part of the product surface for coding agents. That is an inference from the launches, based on both companies highlighting governance, observability, runtime controls and private-data access rather than only model quality or code-generation speed. ### Where does this leave developers using coding agents? GitHub’s public preview means Copilot users can begin testing cloud and local sandboxes now, with the feature documented in GitHub’s Copilot docs and changelog. (blog.cloudflare.com) Cloudflare’s Claude integration is already described in Cloudflare’s engineering post and press materials, which outline runtime options, private service connectivity and control-plane customization for Claude Managed Agents. (github.blog)