OpenAI has updated its agents software development toolkit (SDK), introducing new features designed for businesses to create automated agents running on OpenAI’s models.
The updated SDK incorporates a sandboxing capability, allowing agents to operate in controlled environments. This mitigates the risks associated with unsupervised operation, particularly due to the unpredictable behavior of agents.
Agents can now function within isolated workspaces, accessing files and code only for specific tasks while maintaining system integrity. The SDK also offers an in-distribution harness for frontier models, enabling agents to interact with approved tools and files.
Karan Sharma from OpenAI stated that the launch aims to make the SDK compatible with various sandbox providers. This compatibility is expected to facilitate the development of complex, multi-step “long-horizon” agents.
OpenAI plans further expansion of the Agents SDK, starting with the initial launch of sandbox and harness capabilities in Python, with TypeScript support expected in the future. The company is also developing additional features, including code mode and subagents for both programming languages.
The new SDK capabilities are available to all customers through the API at standard pricing.
