OpenAI adds sandboxing and harness tools to Agents SDK for enterprise AI
Key Points
- OpenAI adds sandboxing to its Agents SDK, isolating agents in controlled workspaces.
- An in‑distribution harness enables agents to run on frontier models while staying confined.
- Initial release supports Python; TypeScript support is planned for later.
- Features aim to let enterprises build longer‑horizon, multi‑step AI agents safely.
- All new tools are available through OpenAI’s standard API pricing.
OpenAI announced a major upgrade to its Agents SDK, introducing sandboxing capabilities and an in-distribution harness for frontier models. The new features let businesses run AI agents in isolated environments and test them with advanced models while keeping systems secure. Initially available in Python, with TypeScript support slated for later, the tools aim to help enterprises build longer‑horizon, multi‑step agents without risking unintended actions. OpenAI says the enhancements will be offered through its standard API pricing.
OpenAI unveiled the latest version of its Agents SDK, a toolkit that lets developers build AI‑driven agents on top of the company’s models. The update adds two core capabilities: sandboxed execution environments and an in‑distribution harness for frontier models. Together, they address a key concern for enterprises—how to give agents the freedom to perform complex, multi‑step tasks while preventing them from accessing or altering unintended parts of a system.
The sandbox feature isolates an agent within a predefined workspace, allowing it to read or write files only in that space. "Running agents in an unsupervised fashion can be risky because they sometimes behave unpredictably," OpenAI explained. By confining the agent, the SDK protects the broader infrastructure from accidental or malicious actions, a safeguard especially important for long‑horizon tasks that involve code generation, data manipulation, or external tool use.
Alongside sandboxing, the SDK now includes an in‑distribution harness that connects agents to frontier models—OpenAI’s most advanced, general‑purpose offerings. In development parlance, a "harness" bundles the supporting components that let a model interact with files, APIs, and approved tools within a workspace. The new harness lets developers deploy and test agents on these cutting‑edge models without leaving the sandbox, streamlining the path from prototype to production.
Karan Sharma, a product manager on OpenAI’s team, told TechCrunch that the launch focuses on compatibility with a range of sandbox providers. "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 said. Sharma added that the combination of sandboxing and harness capabilities should enable users to "build these long‑horizon agents using our harness and with whatever infrastructure they have."
The updated SDK rolls out first for Python developers, with TypeScript support promised in a future release. OpenAI also hinted at upcoming features such as code mode and sub‑agents for both languages, expanding the toolkit’s versatility. All new capabilities will be accessible via the standard OpenAI API, subject to the company’s existing pricing model.
Industry observers see the move as a response to growing demand for enterprise‑grade AI assistants that can operate safely at scale. By giving developers the means to contain agents and leverage the most powerful models, OpenAI positions its platform as a go‑to solution for businesses seeking to automate complex workflows without compromising security.