OpenAI agents SDK

OpenAI has announced a massive update to its Agents SDK, introducing standardized infrastructure designed to move AI agents from prototypes to production. The core of this update is a model-native harness that allows agents to interact with files and commands across a computer, combined with native sandbox execution for secure, isolated operations.

A More Capable Agent Harness

The new SDK harness is built to align with how frontier models—including the newly mentioned GPT-5.4—perform best. Key features include:

  • Standardized Primitives: Integration with the Model Context Protocol (MCP) for tool use, AGENTS.md for custom instructions, and specialized tools for shell execution and code patching.

  • Durable Execution: The harness separates the agent loop from the compute environment. This allows for snapshotting and rehydration, meaning an agent can resume its task from a checkpoint even if its original sandbox container fails or expires.

  • Configurable Memory: Improved orchestration that handles long-horizon tasks across dozens of steps without "context rot."

Native Sandbox Execution

One of the most requested features, native sandboxing, provides agents with a secure workspace to read/write files, install dependencies, and run code safely.

  • Portable Manifests: Developers can use a "Manifest" abstraction to define the agent's workspace, mounting local files or bringing in data from AWS S3, Google Cloud Storage, or Azure Blob Storage.

  • Flexible Providers: The SDK supports built-in integration with sandbox providers like Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel, or developers can bring their own.

Security and Scalability

By separating the harness from the compute layer, OpenAI aims to prevent prompt-injection and data exfiltration attempts. This architecture also enables high scalability: a single agent run can orchestrate multiple sandboxes, route subagents to isolated environments, and parallelize work for faster completion.

Availability and Roadmap

The new Agents SDK capabilities are generally available starting today via the OpenAI API.

  • Language Support: Launching first in Python, with TypeScript support coming soon.

  • Future Features: OpenAI is working on bringing "Code Mode" and "Subagents" (multi-agent orchestration) to both language SDKs in the coming months.

This SDK represents a shift toward "production-viable" agentic AI. It provides a blueprint for how developers can build agents that handle complex clinical records, financial analysis, or automated coding tasks with enterprise-grade security and reliability.