OpenAI Upgrades Agents SDK With Sandboxes and Memory Controls

AI agents have become the next big software pitch, with OpenAI and Anthropic both trying to sell businesses on systems that can do more than answer prompts. The real test, though, is whether those agents can handle actual work once the demo ends. And for that on April 15, OpenAI upgraded its Agents SDK with native sandboxes, tighter memory controls, and a stronger harness for developers building agents that need to work inside real environments.

The sandbox feature ,OpenAI says,  is meant for agents that need a real workspace, not just a prompt window, which means tasks involving files, commands, generated outputs, approval pauses, and resumed work later. The more useful agents become, the more important it is to control where they run and what they remember.

Also Read: Why Allbirds Sold the Brand That Made It Famous and Turned to AI

OpenAI is not just adding another developer feature. The architecture it is showing off separates the trusted harness from the sandbox where untrusted code runs. In practice, that means the agent can work through files, run commands, and write code inside a controlled environment, while secrets, API access, and other sensitive parts stay outside that sandbox.

OpenAI memory controls and Python rollout matter

The memory piece matters too, and it is where the update starts to feel more mature. OpenAI says sandbox memory is separate from normal conversational memory and is designed to carry forward things like user preferences, corrections, project lessons, and task summaries across runs. The system starts with a summary file and only pulls in deeper memory when needed, which is a cleaner answer to the old problem of agents having to relearn the same context over and over.

There is still a catch. The new harness and sandbox capabilities are launching first in Python, with TypeScript support planned for a future release. Even so, the direction is clear.