On Stream: Agent Sandboxes

Why It Matters

Sandboxing AI agents protects critical infrastructure and credentials while allowing developers to harness powerful automation without risking unintended system changes.

Key Takeaways

• Agent sandboxes isolate AI tools from host file and network access.
• Four sandbox categories: native kernel, containers, mini‑VMs, cloud‑proxied.
• Docker SBX provides standalone VM sandbox without Docker Desktop.
• Cloud Code’s built‑in sandbox uses macOS Seatbelt and Linux bubblewrap.
• Sandboxing enables YOLO mode while preventing accidental system changes.

Summary

The video dives into the emerging practice of sandboxing AI‑driven agent harnesses—tools like Claude, Copilot, and other terminal‑based assistants. By confining these agents to isolated environments, developers can control file system and network permissions, preventing rogue actions such as unwanted installations or accidental deletions.

Four primary sandbox models are outlined: native kernel‑level restrictions (Seatbelt on macOS, bubblewrap on Linux), container‑based isolation (Docker), lightweight mini‑VMs like Docker’s new SBX command‑line tool, and emerging cloud‑proxied solutions. The host‑native option is the easiest entry point on macOS/Linux, while Docker SBX offers a standalone VM without needing Docker Desktop. Built‑in features in Cloud Code let users toggle sandbox modes via a simple "/sandbox" command, with auto‑allow or strict deny settings.

The presenter shares a real‑world anecdote where an agent generated video files in a temporary folder and failed to clean up, illustrating the need for sandbox enforcement. Demonstrations show how Cloud Code’s sandbox leverages the OS kernel, how auto‑allow streamlines repetitive permission prompts, and how tools like bash remain the universal gateway for agent actions, making sandbox configuration critical.

For platform engineers and DevOps teams, proper sandboxing balances security with productivity: it enables a “YOLO” mode where agents can operate freely inside a confined space, while safeguarding host credentials, infrastructure keys, and critical directories. As the market matures, these isolation layers will become standard safeguards for AI‑augmented development workflows.

Original Description

Comparing agent harness sandboxes like ohno, docker sandboxes, openshell, veto, and more.

🙌 I've launched the Agentic DevOps Guild, which is my premium community for accelerating your AI adoption for DevOps, CI/CD, platform engineering, and SRE. It includes courses, regular meetups, workshops, and mentorship. 🍾 https://www.bretfisher.com/theguild

🗞️ Sign up for my weekly newsletter for the latest on upcoming guests and what I'm releasing: https://www.bretfisher.com/newsletters/cloud-native-devops