OpenAI updates its Agents SDK to help enterprises build safer, more capable agents

OpenAI has released a significant update to its Agents SDK, introducing sandboxing capabilities that allow enterprise AI agents to operate in controlled, isolated computer environments. This addresses a critical safety concern: running autonomous agents without supervision can be risky due to their occasionally unpredictable nature. With the new sandbox integration, agents can work in siloed workspaces, accessing files and code only for specific operations while protecting the broader system's integrity. The update also includes an in-distribution harness for frontier models like GPT-4, enabling agents to interact with approved tools and files within their designated workspace.

According to OpenAI's Karan Sharma, this launch makes the existing Agents SDK compatible with various sandbox providers, allowing users to build complex 'long-horizon' agents using their preferred infrastructure. These long-horizon tasks involve multi-step, sophisticated workflows that require greater reliability and safety measures. The new capabilities are launching first in Python, with TypeScript support planned for a future release. OpenAI also indicated it's working on additional agent features like code mode and subagents for both programming languages. All customers can access these new SDK capabilities through the standard API with regular pricing, making advanced agent development more accessible to enterprises looking to automate complex business processes safely.