The Next Evolution of the Agents SDK

OpenAI has significantly updated the Agents SDK, providing a standard harness that lets agents more reliably perform file inspection, command execution, code editing, and long-running tasks.

The core is strengthened infrastructure around the model. The new harness integrates frontier agent primitives like MCP, skills, AGENTS.md, shell, and apply patch, letting developers focus more on domain logic.

Native sandbox execution has also been added.

• Agents run in a controlled workspace, allowing them to safely handle files, dependencies, and tool use.
• Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, Vercel, and others are supported by default, and connecting your own sandbox is also possible.
• The Manifest abstraction handles local mounts, output directories, and storage integration with services like S3 / GCS / Azure Blob / R2 in a portable way.

Security and durability are also emphasized. Separating the harness from compute reduces the risk of prompt injection and credential exposure, while snapshotting and rehydration allow execution to resume from a checkpoint even if the container disappears. When needed, multi-container parallelization is also supported by running multiple sandboxes in parallel.

Oscar Health was introduced as a case study. It describes automating, at a more production-grade level than before, a workflow that requires precisely identifying each encounter boundary in complex clinical records.

The new features are provided as GA, with standard pricing based on token and tool usage. Currently it's available first in Python, with TypeScript support, code mode, and subagents to be added later.