Docker is introducing a new way to run coding agents in isolated environments using container-based sandboxes. This approach allows agents to access necessary resources without compromising the local system's safety, addressing security concerns as agents become more autonomous. The current experimental version supports Claude Code and Gemini CLI, with plans for broader agent compatibility.

This article discusses two patterns for connecting agents to isolated execution environments called sandboxes. The first pattern runs the agent inside the sandbox, while the second keeps the agent on a local server and uses the sandbox as a tool. Each method has its own benefits and trade-offs regarding security, update speed, and separation of concerns.