OpenAI updates Agents SDK to improve enterprise safety

OpenAI has released an update to its Agents SDK—a software development toolkit for building AI agents—introducing new features designed to help businesses create agents that run on its models. The release comes as companies like OpenAI and Anthropic compete to provide enterprises with tools to build automated systems.

The update introduces sandboxing, which refers to a controlled computer environment for running software. This feature allows agents to operate in controlled environments, addressing safety concerns in enterprise deployment. Specifically, running agents in a totally unsupervised fashion can be risky due to their occasionally unpredictable nature. By integrating sandboxing, developers can restrict agents to siloed workspaces where they can access files and execute code only for specific operations, protecting the broader system’s integrity.

Additionally, the update provides developers with an in-distribution harness for frontier models, which are advanced, general-purpose AI models. In agent development, the harness refers to the components of an agent besides the model it is running on. This harness allows agents to work with files and approved tools within a workspace, and often allows companies to both deploy and test the agents.

The integration is designed to help developers build long-horizon agents. While standard tasks are typically simpler, long-horizon tasks are generally considered to be more complex, multi-step work.

“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,” said Karan Sharma, who works on OpenAI’s product team. Sharma noted that the goal is to enable developers to build these long-horizon agents using OpenAI’s harness alongside whatever infrastructure they already have.

Why it matters

OpenAI is updating its Agents SDK to include sandboxing and harness capabilities, aiming to help enterprises build safer, more capable agents that run on its models.