OpenAI Expands Agents SDK With Sandboxing and Harness Features

OpenAI has introduced major updates to its Agents SDK, adding sandboxing and harness features to help developers build and deploy safer, more capable AI agents, announced in a press release.

The updated SDK now supports native sandbox execution, allowing agents to work within controlled computer environments. This setup lets agents inspect files, run code, and use tools securely while isolating them from sensitive systems. Developers can use built-in support for sandbox providers such as Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel, or bring their own sandbox solution.

The new harness provides standardized infrastructure that aligns with how OpenAI models operate, improving performance on complex, multi-step tasks. It includes configurable memory, sandbox-aware orchestration, and integrated tools for file handling and code execution. The harness also supports separation of compute and harness components, enhancing security and durability by isolating model-generated code from sensitive credentials.

The updated Agents SDK is available to all customers through the API using standard pricing. The new features launch first in Python, with TypeScript support planned for a future release.