
OpenAI has introduced new capabilities to its Agents software development kit, adding sandboxing and advanced harness tools aimed at helping businesses build and deploy AI agents more safely and effectively. The update focuses on enabling more complex, multi-step automation while addressing risks tied to autonomous systems.
Sandboxing To Control Agent Behavior
The updated SDK includes a sandbox feature that allows AI agents to operate within controlled environments. This setup limits access to specific files, code, and tools required for a given task, while isolating the rest of the system.
The approach is designed to reduce risks associated with agents running unsupervised, particularly given their potential to behave unpredictably in open environments.
Harness Tools For Deployment And Testing
OpenAI has also introduced an in-distribution harness for its frontier models, providing developers with a framework to deploy and test agents within defined workspaces. In this context, the harness refers to the supporting components that manage how an agent interacts with tools, data, and its operating environment.
The company said the harness enables agents to perform tasks using approved resources while maintaining structured oversight.
Support For Complex, Long-Horizon Tasks
According to Karan Sharma, the new features are intended to support “long-horizon” tasks, which involve multi-step processes and extended workflows. The combination of sandboxing and harness tools is expected to help developers build agents capable of handling more advanced operations across enterprise systems.
Compatibility And Development Roadmap
The initial release of these features is available in Python, with support for TypeScript planned for a future update. OpenAI said it is also working to expand agent capabilities, including features such as code mode and subagents, across both programming environments.
Availability And Pricing
The updated Agents SDK is available to all customers through OpenAI’s API and follows the company’s standard pricing model.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.
