Agentic AI has emerged as the latest breakthrough in the technology sector, with firms such as OpenAI and Anthropic competing to provide businesses with the resources required to build these automated assistants. In line with this effort, OpenAI has enhanced its agents software development toolkit (SDK), rolling out several fresh features aimed at enabling companies to develop custom agents powered by OpenAI’s models.

Among the SDK’s new additions is a sandboxing function, which lets agents function within secure, isolated computing settings. This measure addresses the potential hazards of operating agents without oversight, given their sometimes erratic behavior. Through sandbox integration, agents can be confined to a designated workspace, interacting solely with relevant files and code for specific tasks, thereby safeguarding the broader system’s stability.

Additionally, the updated SDK includes an in-distribution harness for frontier models, enabling agents to handle files and authorized tools inside a workspace. In agent development, a “harness” encompasses all non-model components of an agent. An in-distribution harness typically supports both deployment and testing of agents built on frontier models—the most sophisticated, versatile models currently available. According to the company, combining this with the new harness functionalities should empower users “to go build these long-horizon agents using our harness and with whatever infrastructure they have.” These “long-horizon” tasks usually involve intricate, multi-stage operations.

OpenAI has indicated that it will keep expanding the Agents SDK moving forward. For now, the new harness and sandbox features are debuting in Python, with TypeScript compatibility slated for a future update. The company is also developing additional agent capabilities, like code mode and subagents, for both Python and TypeScript.

These enhanced Agents SDK tools are accessible to all customers through the API and will follow standard pricing.