How to Actually Run Your Coding Agent Safely (And Avoid the Horror Stories)

The video begins by emphasizing the security risks of running AI coding agents directly on a user's machine, especially with powerful commands like `rm -rf`. It introduces Docker Sandbox as a solution for running these agents in isolated environments. The presenter walks through the setup process for Docker Sandbox on both macOS and Windows, providing the necessary commands. The core of the video explains the concept of isolation layers within Docker Sandboxes, detailing how different layers protect the host system from the agent's actions. The presenter discusses the inherent risks of un-sandboxed agents, such as the ability to access and delete files, run destructive commands, or interact with databases and network services inappropriately. The video showcases real-world examples of these risks from GitHub issues, including data loss and home directory deletion. Finally, it demonstrates how to run an agent within a Docker Sandbox using the `sbx run <agent_name>` command, ensuring that the agent's operations are contained and do not affect the host system.