Agentic AI represents the latest triumph in the tech sector, with firms such as OpenAI and Anthropic competing to equip businesses with the resources necessary to develop these automated assistants. In this pursuit, OpenAI has updated its agent software development toolkit (SDK), unveiling several new features aimed at assisting enterprises in crafting their own agents that operate on OpenAI’s models.

Among the SDK’s latest features is a sandboxing function that enables agents to function within controlled computing environments. This is crucial as deploying agents without oversight can pose risks due to their occasionally erratic behavior.

Thanks to the sandbox feature, agents can function in a contained manner within a specific workspace, accessing files and code solely for designated tasks, all while preserving the overall system integrity.

Additionally, the updated SDK offers developers a harness tailored for frontier models, allowing agents to handle files and authorized tools within a workspace, as stated by the company. (In the context of agent development, a “harness” refers to the components of an agent apart from the model it operates on. A harness geared towards in-distribution typically enables firms to deploy and test agents utilizing frontier models, which are seen as the most advanced, versatile models available.)

“This launch fundamentally aims to adapt our existing Agents SDK to ensure compatibility with numerous sandbox providers,” remarked Karan Sharma, a member of OpenAI’s product team, to TechCrunch.

The goal is that this, combined with the new harness features, will empower users “to build these long-horizon agents using our harness and the infrastructure they possess,” he explained.

Tasks classified as “long-horizon” are typically viewed as more intricate and multi-step assignments.

OpenAI indicated its intention to progressively enhance the Agents SDK, but initially, the new harness and sandbox features will debut in Python, with TypeScript support anticipated for a future update. The firm also stated it is working on adding additional agent capabilities such as code mode and subagents to both Python and TypeScript.

The new capabilities of the Agents SDK are made available to all clients through the API, adhering to standard pricing.