OpenAI has expanded its Agents Software Development Kit (SDK) with new features aimed at enabling businesses to build autonomous AI agents that operate within controlled environments. The update introduces sandboxing capabilities, allowing agents to function in isolated workspaces where they can access specific files and tools without compromising broader system integrity. This addresses growing concerns over the unpredictable behaviour of autonomous systems when deployed without oversight.
The revised SDK also includes an in-distribution harness for frontier models, enabling agents to interact with approved resources in a workspace while supporting deployment and testing of advanced, general-purpose AI models. In agent development terminology, the harness refers to the supporting infrastructure beyond the model itself, and its inclusion here allows for more secure and flexible integration with existing enterprise systems.
Karan Sharma, a member of OpenAI’s product team, explained that the update makes the Agents SDK compatible with a range of sandbox providers, allowing users to construct long-horizon agents—complex, multi-step processes—using their preferred infrastructure. This broadens the potential applications for enterprises seeking to automate sophisticated workflows.
Initially available in Python, with TypeScript support planned, the new capabilities are accessible to all customers via the API under standard pricing. OpenAI has signalled further enhancements, including expanded support for features such as code mode and subagents, across both programming languages.
The move positions OpenAI competitively in a rapidly evolving market, where rivals like Anthropic are also advancing agent-building tools. By prioritising security, flexibility, and compatibility, the company aims to make autonomous agents more viable for enterprise adoption, while mitigating the risks associated with unsupervised AI operations.
