OpenAI Agents SDK Update Enhances Enterprise AI Safety
▼ Summary
– OpenAI has updated its Agents SDK to help businesses create automated agents using its models.
– A key new feature is sandboxing, which lets agents operate in controlled, isolated environments to reduce risk.
– The update also introduces an in-distribution harness, allowing agents to work with files and approved tools within a workspace.
– These capabilities aim to enable the development of “long-horizon” agents capable of complex, multi-step tasks.
– The new features are initially launching in Python for all API customers, with TypeScript support and more capabilities planned for the future.
The race to build and deploy agentic AI is intensifying, with major players releasing tools designed for enterprise adoption. OpenAI has rolled out a significant update to its Agents SDK, providing developers with new features focused on safety and control. This move aims to equip businesses with the infrastructure needed to create reliable, automated assistants powered by the company’s advanced models.
A central component of this update is the introduction of sandboxing capabilities. This feature allows developers to run AI agents within isolated, controlled computer environments. Given that agents can sometimes behave in unpredictable ways, operating them without supervision poses inherent risks. Sandboxing mitigates this by letting an agent work in a siloed workspace, where it can access specific files and code only for authorized tasks, thereby safeguarding the broader system’s integrity.
Complementing the sandbox is a new in-distribution harness for frontier models. In agent development, a “harness” refers to the surrounding framework that manages an agent’s operations beyond its core AI model. This specific harness enables agents to interact securely with files and approved tools within a designated workspace. It also facilitates both the deployment and rigorous testing of agents built on frontier models, which are the most advanced and general-purpose AI systems available today.
According to Karan Sharma of OpenAI’s product team, the core objective of this launch is to enhance compatibility. “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 stated. The combination of sandboxing and the new harness is intended to empower users to construct long-horizon agents. These are systems capable of executing complex, multi-step workflows using their preferred infrastructure.
Initially, these new safety and control features are launching for the Python programming language, with support for TypeScript scheduled for a future update. OpenAI plans to continue expanding the SDK’s functionality, with ambitions to introduce additional agent capabilities such as code mode and subagents across both language platforms. All customers can access the updated Agents SDK through the standard API, which will apply regular pricing.
(Source: TechCrunch)
