Docker ships microVM sandboxes to run AI coding agents in YOLO mode safely

Docker has launched Sandboxes, a tool that runs AI coding agents inside disposable microVM environments isolated from the host machine. Each sandbox enforces user-defined network and filesystem controls while providing a hard security boundary, and it spins up and tears down faster than a conventional VM. Because the isolation is real rather than process-level, agents can even launch their own Docker containers inside the sandbox and operate unattended in a full development environment. Installation is a brew one-liner and does not require Docker Desktop.

The pitch targets a specific pain point: coding agents increasingly run with autonomy flags like —dangerously-skip-permissions (“YOLO mode”), which skip approval prompts for speed but let an agent touch anything on the machine. Docker’s answer is to contain that blast radius per-agent in a dedicated microVM rather than reining in the agent itself. A single sandbox setup works across the major CLI agents, including Claude Code, Gemini CLI, GitHub Copilot CLI, Codex, Kiro, and OpenCode, with a path to define custom ones.

Sandboxes cover the local, per-developer case; centralized controls are gated behind a separate paid tier called Docker AI Governance, which pushes network access policies, filesystem rules, and org-wide MCP governance to every developer’s machine from one definition. The split reflects where Docker sees monetizable value as agentic development scales — free isolation for individuals, enforced policy for teams.