OpenAI has updated its agent software development kit (SDK) that supports companies developing AI agents, TechCrunch reported on April 15 local time.

The key change in the update is the addition of a sandbox feature that isolates agents so they operate only within a controlled computer environment.

The sandbox feature helps protect overall system integrity while allowing access only to specific files and code. It is intended to block unpredictable behavior that could occur if an agent runs without full oversight.

An in-distribution harness feature was also added. In agent development, a harness refers to other components that make up an agent besides the model.

Karan Sharma (카란 샤르마) of OpenAI's product team said, "The core of this release is making the existing agent SDK compatible with all sandbox providers," adding, "Combined with harness capabilities, it enables users to build agents that handle long-horizon and complex ('long-horizon) tasks with the infrastructure they have."

Long-horizon and complex tasks refer to complicated work carried out over multiple steps. The new features are available first in Python, and TypeScript support will be added later. OpenAI is also developing additional agent features such as code mode and sub-agents.