OpenAI has updated its Agents Software Development Toolkit (SDK), adding sandboxing capabilities and in-distribution harnesses. These features enable enterprises to build safer, more capable automated agents with controlled environments, enhancing security while allowing access to specific files and code for particular operations.

The new sandbox integration is crucial as it limits the operational scope of AI agents, reducing the risk of unintended consequences from their unpredictability. Developers can now use this feature within a workspace, ensuring that agents only interact with approved tools and resources, thus maintaining system integrity.

OpenAI also introduced an in-distribution harness for frontier models, which allows companies to deploy and test advanced AI agents running on the most sophisticated general-purpose models available. This tool can help businesses tackle complex, multi-step tasks more effectively. The SDK’s new capabilities are initially launching with Python support, followed by TypeScript.

With these enhancements, OpenAI aims to empower users to build long-horizon agents using their preferred infrastructure and testing environments. This update is part of a broader strategy to expand the Agents SDK over time, focusing on adding features like code mode and subagents for both Python and TypeScript users.