Docker Sandboxes: Safe MicroVM Runs for AI Coding Agents

Docker launched Sandboxes on August 10: isolated microVM environments that let AI coding agents run without user supervision. The product is the container giant’s first dedicated answer to the biggest operational problem of agentic coding — giving autonomous agents a place to work that can’t damage the host or wander the network.

Out of the box it supports the major coding agents — Claude Code, Gemini CLI, Copilot CLI, Codex, OpenCode and Kiro — with filesystem and network controls and, for the risk-tolerant, a --dangerously-skip-permissions YOLO mode.

Key facts

• Disposable microVMs, not containers: hardware-isolated execution that can be thrown away after the run — the agent’s environment is the artifact, not the host.
• Agent support: Claude Code, Gemini CLI, Copilot CLI, Codex, OpenCode, Kiro — the leading harnesses, day zero.
• Controls: filesystem access limits and network controls per environment; permission gates per tool call.
• YOLO mode: --dangerously-skip-permissions — intentional, documented, and named as a risk: unsupervised agents run with full permissions.
• Why now: agent harnesses have been the fastest-growing consumer of CI-style compute; sandboxing them is the professionalization step between “demo on my laptop” and “production agent fleet.”

Why it matters

• The sandbox becomes the new container. Docker defined the compute primitive of the last decade; Sandboxes defines the compute primitive of agent execution — isolated, disposable, per-task.
• Safety is the adoption gate for unsupervised agents. Every serious deployment of coding agents needs containment; Docker just made it a commodity feature rather than an in-house build.
• The YOLO flag is the market’s honesty check. Exposing the permission-skip option puts the safety choice on record — and guarantees a future incident report that names it.

What to watch

• Default agent-harness integration depth: does Claude Code / Codex make Sandboxes a one-flag default or a YOLO-mode footnote?
• Cost per run vs. bare-metal containers — microVM overhead matters for high-frequency agent loops.
• Security research on Sandboxes escape surfaces — microVMs are the newest favorite target in agent-security research.

Official source

• Docker: Docker Sandboxes product page
• Coverage: AI Weekly — top AI stories, August 10, 2026

Updated August 10, 2026.