OpenAI has expanded its Agents SDK with a harness for agents that work across files and tools, plus native sandbox execution. The initial release is for Python, with TypeScript support planned later.
The architecture separates orchestration from the environment where model-generated code runs. That lets credentials remain outside the sandbox, while agent state can be snapshotted and restored if an execution environment fails.
This is a sign of the agent stack maturing. The competitive surface is moving beyond models and prompts into runtimes, isolation, observability, and the mechanics of work that lasts longer than one API call.