Docker Sandboxes offer disposable microVM environments for AI agents

Docker has launched Docker Sandboxes, a new product that gives AI coding agents disposable, isolated environments to run in. Built on microVM isolation, each sandbox gives an agent its own dedicated environment with the developer's project workspace mounted in, so agents can install packages, modify configs, and even spin up their own Docker containers without touching the host machine.

The product is designed for the wave of coding agents that need safe, unattended execution: Claude Code, Gemini CLI, Copilot CLI, Codex, OpenCode, and Kiro are all supported. Sandboxes are disposable by default, spin up faster than traditional VMs, and tear down in a single command. Docker says agents can work without manual review, permission prompts, or supervision, while customizable network and filesystem controls set the boundaries.

Docker also ties the product into its AI Governance capabilities, letting teams enforce those controls org-wide. The CLI installs in seconds on macOS via Homebrew or on Windows via winget, and Docker is offering a free tier to get started. The launch positions Docker to become the default runtime for agentic coding workflows, where autonomous agents increasingly need safe execution environments that keep host machines protected.