Looking at the agent ecosystem lately, it’s clear that we're moving beyond simple API integration toward sandbox technology, which allows for secure, real-time code execution within an app while maintaining safety. It's incredibly hot right now! OpenAI's newly unveiled ChatGPT Work and Canva's Canva Code 2.0 are perfect examples of this trend.

First, ChatGPT Work uses Codex technology within its desktop app to safely run multi-step automation. It offers granular security options like sandbox_mode and task approval rules like approval_policy, which limit the scope of file changes. For needs requiring stronger control and network isolation, it recommends connecting via a Model Context Protocol (MCP) gateway linked to a Docker container. It’s a smart approach that gives agents powerful capabilities while keeping the local host computer safe.

Here is an example of a ChatGPT Work configuration file for specifying a safely restricted workspace.

{
  "sandbox_mode": "workspace-write",
  "approval_policy": "on-request"
}

Canva's simultaneously released Canva Code 2.0 is also quite interesting! Powered by Anthropic's Claude, it processes responsive compilation—such as for websites or mini-apps—75% faster. The generated code runs client-side within a secure iframe sandbox in the browser. Furthermore, Canva’s design model parses this code, allowing users to perform visual editing at the layer level.

Beyond AI that simply provides good answers, it seems that AI equipped with reliable and securely isolated execution environments is becoming the real competitive edge for enterprise agents. I’m really looking forward to seeing how developers will leverage this in-app sandbox security moving forward!