OpenAI Upgrades Agents SDK with Sandbox and Harness Tools to Boost Enterprise AI Agent Safety

OpenAI has rolled out a major update to its Agents Software Development Kit (SDK), aiming to help businesses build more capable and safer AI agents. The update comes as competition in the “agentic AI” space intensifies, with firms like OpenAI and Anthropic pushing tools that allow companies to automate increasingly complex digital tasks using advanced models.

A key feature in the new release is sandboxing, which allows AI agents to operate within controlled environments. This means agents can access specific files, code, and tools within a restricted workspace without putting broader systems at risk. The feature is designed to reduce errors and improve safety, especially given that AI agents can sometimes behave unpredictably when left unsupervised.

OpenAI also introduced what it calls an “in-distribution harness,” a framework that lets developers deploy and test agents using frontier AI models while giving them structured access to tools and files. In simple terms, the harness acts as the supporting system around the AI model, helping it function effectively in real-world enterprise environments.

According to OpenAI product team member Karan Sharma, the goal is to make it easier for developers to build “long-horizon agents” that can handle multi-step, more complex tasks. These upgraded capabilities are designed to integrate with various infrastructure setups, giving enterprises flexibility while still maintaining control and safety.

The updated Agents SDK is now available via API for all customers under standard pricing. Initially, the new sandboxing and harness features are launching in Python, with TypeScript support expected in future releases. OpenAI also confirmed it is working on expanding the SDK further, including additional agent capabilities such as code mode and subagents.

source: Techcrunch