Modern AI Agents are coupled with runtime sandbox for filesystem access and runtime isolation. Sometimes referred to as Computer-Use Agents (CUA), they autonomously code, use terminal, take notes, and access Internet like human does. In this post, let’s dive into agent design principles beneath popular scaffolds like Claude Code, Gemini (Google AI Studio) and ChatGPT Agent, and unveil how agents benefit from better context management, broader action space and autonomous workflows.

Why Sandbox?

Claude Code – More than Just Code

MiniMax Agent – The “Deepest”

Other CUA

  • ChatGPT Atlas/Agent Mode
  • Google AI Studio

Agentic RL – Rollout Env

  • OpenEnv
  • OpenHands