OpenAI has released a significant update to its Agents software development kit (SDK), introducing new sandboxing and harness features designed to help businesses build safer and more capable AI agents. The announcement, made on Tuesday, targets enterprise customers looking to deploy automated assistants powered by OpenAI's frontier models for complex, long-horizon tasks.

The core of the update is a new sandboxing ability, which allows AI agents to operate within controlled, isolated computer environments. This addresses a critical safety concern, as agents running unsupervised can sometimes behave unpredictably. With this integration, agents can access files and execute code for specific operations within a designated workspace without compromising the broader system's integrity.

New Harness for Frontier Models

Alongside sandboxing, OpenAI is providing an "in-distribution harness" for its most advanced models. In agent development, a harness refers to the supporting components around the core AI model. This new harness will allow agents to interact safely with files and approved tools within a workspace, facilitating both deployment and testing.

Karan Sharma of OpenAI's product team explained the rationale to TechCrunch: “This launch, at its core, is about taking our existing agents SDK and making it so it’s compatible with all of these sandbox providers.” He added that the goal is to enable users “to go build these long-horizon agents using our harness and with whatever infrastructure they have.”

Initial Rollout and Future Plans

The new sandbox and harness capabilities are launching first for the Python programming language, with support for TypeScript planned for a future release. All customers can access the updated SDK through the standard API, and it will use OpenAI's existing pricing model.

The company confirmed it is continuing to develop the Agents SDK and is working to bring additional features—such as code mode and subagents—to both Python and TypeScript environments. This move places OpenAI in direct competition with other AI firms like Anthropic, which are also racing to provide enterprise-grade agent-building tools.

The Drive for Safer Agentic AI

The update reflects a growing industry focus on making "agentic AI"—where AI systems can autonomously perform multi-step tasks—more reliable for business use. Sandboxing is seen as a fundamental step to prevent agents from taking unintended actions that could disrupt operations or access sensitive data.

By offering these tools, OpenAI aims to lower the barrier for enterprises to adopt and experiment with advanced AI assistants for complex workflows, while providing built-in safeguards to manage the associated risks.