Agentic AI is the tech industry’s newest success story, and companies like OpenAI and Anthropic are racing to give enterprises the tools they need to create automated helpers. OpenAI is now leaning further into that push, updating its agents software development toolkit (SDK) with new features aimed at helping businesses build agents that run on the backs of OpenAI’s models.
The headline addition is sandboxing—basically, a way for agents to operate inside controlled computer environments. The point is pretty straightforward: running agents entirely unsupervised can be risky, because their behavior can be a little unpredictable when they’re left to roam.
With the sandbox integration, agents can work in a siloed capacity within a particular workspace. They can access files and code only for particular operations, while the rest of the system is kept protected. Or, to put it more plainly, it’s like giving an agent a room to work in and a set of rules for the door. In practice, that matters when enterprises don’t want experiments spilling into the rest of their infrastructure—there’s usually enough chaos already, with meetings and access permissions and all.
Alongside sandboxing, the updated SDK also adds an in-distribution harness for frontier models. Misryoum newsroom reporting describes this as a way for agents to work with files and approved tools within a workspace. The “harness” part is important too: it refers to the agent components beyond the model itself, and an in-distribution setup typically supports both deploying and testing agents. Frontier models, as Misryoum editorial desk noted, are considered the most advanced, general-purpose models available, so having a controlled harness for them is the kind of thing developers usually ask for after the first prototype works… sort of.
Karan Sharma, who works on OpenAI’s product team, said the launch is “about taking our existing agents SDK and making it so it’s compatible with all of these sandbox providers.” Misryoum newsroom reported the goal is that, paired with the new harness capabilities, users can build long-horizon agents using the harness with whatever infrastructure they already have.
Long-horizon tasks are generally more complex and multi-step work. That’s also where agent systems can get tricky: more steps mean more places for something unexpected to happen. So it’s not just about capability—it’s about giving teams a way to test and constrain the behavior as the agent tries to complete longer objectives.
OpenAI says the company will continue expanding the Agents SDK over time. Initially, though, the new harness and sandbox capabilities are launching first in Python, with TypeScript support planned for a later release. The company also said it’s working to bring more agent capabilities like code mode and subagents to both Python and TypeScript. Misryoum editorial team noted that these new Agents SDK capabilities are being offered to all customers via the API, using standard pricing. And somewhere in an office, someone’s coffee is going cold while they figure out whether their agent can now be safely unleashed for the next test run—at least, that’s the vibe you get when a tool like this shows up.
Ticketmaster ruled an illegal monopoly in Manhattan jury verdict