A practical guide to AI agent sandboxing: isolated environments, scoped tools, fake side effects, approval gates, and the containment patterns that let you test agents safely before production access.