The Next Evolution of the Agents SDK

The Agents SDK has expanded its standard infrastructure for agents. At its core are a model-native harness that lets models move between files and tools, and a native sandbox that executes tasks safely.

The new harness has been strengthened for agents that work with documents, files, and systems. It provides configurable memory, sandbox-aware orchestration, and Codex-like filesystem tools, and it also standardizes and absorbs frontier agent patterns such as MCP, skills, AGENTS.md, shell tools, and apply patch tools.

Developers can give their agents a controlled workspace and explicit instructions, and have them read the necessary files and verify grounding. OpenAI explains that this approach is closer to how models naturally execute long-running tasks or complex work that spans multiple tools, and that it improves reliability and performance.

The native sandbox safely handles reading and writing files, installing dependencies, and executing code. Users can configure their own sandbox, or take advantage of support for Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel.

For environment portability, a new Manifest abstraction has also been added. It lets you handle local file mounting, defining output directories, and loading data from storage such as AWS S3, Google Cloud Storage, Azure Blob Storage, and Cloudflare R2 in a single, unified way.

On the security and resilience side, separating the harness from compute is key. Credentials are isolated with prompt injection and exfiltration attempts in mind, and snapshotting and rehydration allow work to resume from the last checkpoint even if a container dies.

Scalability was also emphasized. You can use one or multiple sandboxes, spin up a sandbox only when needed, send subagents to isolated environments, and run in parallel across containers to work faster.

Pricing is generally available and follows the API's standard pricing structure. It is currently available first in Python, with TypeScript support planned for later. After that, plans call for adding code mode and subagents to both languages.