OpenAI has released an updated version of its Agents SDK, introducing sandbox and harness capabilities aimed at helping enterprises build autonomous agents with greater safety controls and functionality.
The sandbox feature is the centerpiece of this release. It allows agents to operate within controlled, isolated computer environments rather than running unsupervised—a safeguard that OpenAI says is necessary because agents can behave unpredictably. Within the sandbox, agents work in siloed workspaces with restricted access to files and code, accessing only what is needed for specific operations while maintaining overall system integrity.
Alongside sandboxing, the update introduces an in-distribution harness for frontier models. In agent development, the harness refers to all components of an agent apart from the underlying model itself. OpenAI's in-distribution harness enables developers to deploy and test agents running on frontier models, which the company classifies as its most advanced, general-purpose models. This harness allows agents to work with files and approved tools within a workspace.
Karan Sharma, who works on OpenAI's product team, explained to TechCrunch that the core objective is making the SDK compatible with multiple sandbox providers. "The hope is that this, paired with the new harness capabilities, will allow users to go build these long-horizon agents using our harness and with whatever infrastructure they have," Sharma said. Long-horizon tasks are generally considered to be complex, multi-step work requiring coordination across multiple operations.
The initial rollout prioritizes Python compatibility, with TypeScript support arriving in a later release. OpenAI also indicated it is developing additional agent capabilities, including code mode and subagents, for both Python and TypeScript. All new Agents SDK capabilities are available to API customers at standard pricing.
OpenAI said it will continue expanding the Agents SDK over time, but did not specify a timeline for additional features.