OpenAI updates Agents SDK with built-in sandbox and native harness for long-running agents

OpenAI has updated the Agents SDK, adding native sandbox execution and a model-native harness, an architectural layer tightly integrated with the capabilities of a specific model. It lets developers build long-running agents that work safely with files, tools and external systems.

Before this update, building agents on OpenAI's stack meant implementing isolation and state management yourself. The SDK offered tools for function calling and handling streams, but complex scenarios required significant engineering effort from teams: an agent that reads and modifies files, runs code and calls external APIs without creating security threats.

Sandbox execution is the key addition. An agent can now run arbitrary code in an isolated environment without direct access to the host system. This removes one of the main barriers to deploying agents in production: the fear that the model will do something destructive to the file system or the network. Isolation is built in at the SDK level rather than being an external add-on that every team assembles from scratch.

The model-native harness is a layer that knows the model's capabilities and uses them directly. Instead of wrapping the model in a generic interface, the harness adapts to how the model actually handles tasks, tools and context. In essence, it is an attempt to close the gap between what the model was trained to do and how a developer applies it.

In practice, this means agents should be able to run more reliably and for longer. Most agent systems today are limited to a few steps: the context window fills up, threads get lost and errors accumulate. The SDK's new architecture is designed for tasks that take minutes or hours, such as processing large codebases, automating workflows with many tools, and research agents that need to work iteratively with documents.

The update fits into OpenAI's broader strategy of turning agent capabilities into a production standard. The company previously introduced Operator and Computer Use, tools for controlling a browser and an operating system. The Agents SDK is the answer for developers who want a programmable agent engine they can embed in their own product, rather than a ready-made UX solution.

For the developer community, this signals that OpenAI's agent ecosystem is maturing. A native sandbox means the company treats security not as an optional component but as part of the SDK's core infrastructure. That matters: many enterprise customers have put off adopting agents precisely because of security concerns and the unpredictability of model behavior in production.

How this affects competitors is an open question. Anthropic is building similar functionality through its own Agent SDK, and Google is pushing Agent Space on top of Vertex AI. But OpenAI has an edge in the number of developers already working with its API: every SDK improvement immediately reaches a broad audience of early adopters.

More details in