OpenAI has released a major update to its Agents SDK that adds native sandbox execution and a more capable model-native harness, aiming to move the framework from prototype-friendly to production-viable for enterprise workloads. The new capabilities are generally available to all customers via the API at standard token and tool-use pricing, with Python support launching first and TypeScript planned for a future release.
The update addresses a core gap the company identified: developers need more than powerful models to build useful agents — they need systems that support how agents inspect files, run commands, write code, and keep working across many steps reliably. Oscar Health, an early tester, said the updated SDK made it production-viable to automate a critical clinical records workflow that previous approaches couldn't handle reliably, specifically citing the SDK's ability to understand encounter boundaries in long, complex records.
What's new
The harness — the orchestration layer around the model — now includes configurable memory for persisting context across long-running tasks, sandbox-aware orchestration for coordinating multi-step work, Codex-like filesystem tools for reading, writing, and navigating files with model-native operations, and standardized integrations with primitives common in frontier agent systems. The design principle is to align execution with how frontier models perform best, keeping agents closer to the model's natural operating pattern to improve reliability on complex, multi-tool tasks.
Native sandbox execution lets agents run in controlled computer environments with explicit files, tools, and dependencies. Developers can bring their own sandbox or use built-in support for Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel. A new Manifest abstraction makes those environments portable across providers: developers can mount local files, define output directories, and pull data from AWS S3, Google Cloud Storage, Azure Blob Storage, and Cloudflare R2. This gives the model a predictable workspace — where to find inputs, where to write outputs, and how to keep work organized across long-running tasks.
Why it matters
Separating the harness from compute improves security, durability, and scale. By keeping credentials out of environments where model-generated code executes, the architecture reduces prompt-injection and exfiltration risk. Externalizing agent state enables durable execution: if a sandbox container fails or expires, built-in snapshotting and rehydration can restore the agent's state in a fresh container and continue from the last checkpoint. The design also supports scaling — agent runs can use one sandbox or many, invoke sandboxes only when needed, route subagents to isolated environments, and parallelize work across containers.
Our take
The sandbox-native harness is the most concrete step yet toward making OpenAI's SDK a default choice for teams that want to ship agents without stitching together custom infrastructure. The Manifest abstraction and provider-agnostic sandbox layer are practical wins; the real test will be whether the Python-first ergonomics translate cleanly to TypeScript and whether the evaluation tooling keeps pace as agents move from demos to regulated workflows.
Series: Agents SDK orchestration · sandbox update (Aug 2026) · OpenAI Product Stack
