The next evolution of the Agents SDK

OpenAI just pushed a meaningful update to its Agents SDK, and the headline addition is native sandbox execution. Instead of developers bolting on their own isolation layers to keep an agent from doing something destructive while it pokes around files or calls external tools, that containment now ships as part of the SDK itself. It's a plumbing change, but plumbing changes are often the ones that quietly unlock a lot of new behavior.

The second piece is what OpenAI calls a model-native harness. The idea is to give the underlying model a more direct, structured way to manage its own tool calls and file operations rather than relying entirely on external orchestration code that developers write and rewrite for every project. That matters because a lot of the fragility in agent systems today comes from the seams between the model and the scaffolding around it, not from the model's reasoning itself.

Taken together, these two changes are aimed squarely at long-running agents, the kind that might spend minutes or hours working through a multi-step task instead of answering a single prompt and stopping. Long-running work is where sandboxing stops being a nice-to-have and starts being the thing standing between a useful agent and one that accidentally deletes a directory it shouldn't have touched. OpenAI is betting that as more companies try to deploy agents for real operational work, this kind of safety net becomes table stakes rather than a differentiator.

This also fits a pattern OpenAI has followed all year: shipping infrastructure updates that make its own SDK stickier for developers who might otherwise stitch together LangChain, custom orchestration layers, or open-source alternatives. A native sandbox and harness reduce the amount of glue code teams need to write, which is exactly the kind of friction reduction that keeps developers inside one ecosystem instead of shopping around.