OpenAI rolled out new features for its Agents SDK, giving developers a model‑native harness that lets agents move seamlessly between files and tools on a computer. The update adds built‑in sandbox execution, so agents can run tasks safely inside isolated environments. By standardizing the infrastructure, the SDK lowers the barrier for teams moving from prototypes to production in today's development.

Customers who tested the new SDK reported that it now supports configurable memory, sandbox‑aware orchestration, and Codex‑like filesystem tools. The harness also provides standardized integrations with the primitives that frontier agent systems rely on, allowing developers to align execution with the model’s natural pattern. This alignment improves reliability and performance on long‑running or multi‑tool workflows for complex real‑world tasks today.

The SDK adds a Manifest abstraction that lets teams mount local files, set output directories, and pull data from cloud stores such as AWS S3, Google Cloud Storage, Azure Blob, and Cloudflare R2. Built‑in snapshotting and rehydration support durable execution, while multiple sandboxes or isolated sub‑agents scale work across containers. Availability starts in Python, with TypeScript to follow for developers worldwide.