How I Sandbox my AI Agents

AI is taking over world of software development, but it sure ain’t going to take over my computer! I’ve been waist deep in coding with AI, testing out different models, agents, and workflows, but the one constant has been the ability to safely isolate these AI agents in a sandbox to prevent them from going rogue and doing things it shouldn’t (like leaking my api keys or taking down a production database). This is especially important for me since I typically run agents in --dangerously-skip-permissions (aka --yolo) mode, as who wants to babysit them and constantly approve commands they want to run? Although docker now has a dedicated sandbox tool, I personally find it a bit clunky to use compared to non-containerized tools. I also don’t like how each workspace gets its own image, because it means that the development set up needs to be done repetitively for each project (e.g. even if they’re all python-based). In general, I find docker to be great for deployments, but less so for…