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

In the rapidly evolving landscape of Agentic AI, OpenAI is stepping up to equip businesses with advanced tools to create automated assistants. The company has rolled out an updated version of its agents software development toolkit (SDK), which introduces several innovative features aimed at enhancing the safety and efficiency of enterprise applications. One of the standout additions is the sandboxing capability. This feature allows agents to function within controlled computer environments, mitigating the risks associated with their unpredictable behaviors when deployed unsupervised. With sandboxing, agents can operate in isolation, accessing only specific files and code needed for their tasks, thereby safeguarding the overall integrity of the system. Additionally, the new SDK version includes an in-distribution harness for frontier models, enabling agents to interact with files and approved tools in a designated workspace. This harness is crucial for agent development as it encompasses all components necessary for an agent's functionality beyond just the underlying model. It empowers companies to deploy and test their agents using the most sophisticated, general-purpose models available. Karan Sharma, a member of OpenAI's product team, emphasized that this launch focuses on ensuring compatibility with various sandbox providers. The goal is to empower users to develop complex, long-horizon agents that can tackle multi-step tasks effectively. Initially, these new features will be available in Python, with plans to introduce TypeScript support in future updates. OpenAI is committed to continually enhancing the Agents SDK, with more capabilities, such as code mode and subagents, expected to be integrated into both programming languages. The updated SDK is accessible to all customers through the API, maintaining standard pricing structures.