Technology

OpenAI upgrades Agents SDK for enterprise sandboxes

New harness tools gate agent access to files and approved actions, safer autonomy becomes a platform lock-in layer

Images

ScreenshotImage Credits:OpenAI ScreenshotImage Credits:OpenAI Image Credits:OpenAI

OpenAI has updated its Agents SDK, adding sandboxing and new “harness” tooling aimed at enterprises building more autonomous AI agents, according to TechCrunch. The company says the features are available via its API at standard pricing, with the initial release focused on Python and TypeScript support planned later.

The new sandbox integration is a direct response to the problem businesses run into as agents move from chat-style assistance to “long-horizon” work: once a model can write code, read files, and execute multi-step tasks, the failure modes stop being theoretical. A sandboxed environment limits what an agent can touch and where it can run, reducing the blast radius when an agent behaves unexpectedly or is prompted into doing something unsafe. OpenAI’s “in-distribution harness” is meant to make those guardrails practical in production by defining which tools and files an agent is allowed to use inside a workspace—effectively turning agent deployment into a permissions and workflow problem as much as a model-quality problem.

This also shifts the commercial center of gravity. As agents become more capable, the differentiator is less the model and more the surrounding system: the runtime, the tool permissions, the audit trail, and the ability to test and constrain behavior before it reaches internal systems. Enterprises that previously treated AI as a low-risk add-on now face the prospect of software that acts—booking, editing, executing—inside their infrastructure. That creates demand for controlled execution environments, standardized “approved tools,” and repeatable testing setups, which in turn makes vendors’ SDK choices harder to reverse later.

OpenAI’s decision to ship these capabilities through its own SDK—rather than leaving safety and orchestration entirely to third parties—also strengthens its position in the agent stack. If companies build their internal automations around OpenAI’s harness conventions and sandbox providers, switching to a competing model is no longer just an API call change; it becomes a migration of workflows, policies, and operational tooling.

OpenAI says it will expand the Agents SDK over time, including bringing more capabilities such as “code mode” and subagents to Python and TypeScript. For now, the company is selling a safer way to let an AI touch your files and tools without giving it the keys to everything.