OpenAI Agents SDK Update

OpenAI Agents SDK Update

OpenAI has introduced new capabilities to the Agents SDK, providing developers with a standardized, model-native harness and native sandbox execution. This update allows agents to operate across files and tools on a computer safely and reliably, reducing the need for custom infrastructure when moving from prototype to production.

Model-Native Harness for Agent Loops

The Agents SDK now features a more capable harness designed to align execution with the natural operating patterns of frontier models. This alignment improves reliability and performance, particularly for long-running tasks that require coordination across diverse tools and systems.

Key technical components of the updated harness include:

  • Configurable memory: Enhances the agent's ability to maintain state across steps.
  • Sandbox-aware orchestration: Ensures the agent can manage its environment effectively.
  • Codex-like filesystem tools: Provides standardized tools for interacting with documents and files.
  • Standardized integrations: Incorporates primitives common in frontier agent systems.

Because the harness is designed to be turnkey yet flexible, developers can adapt tool use, memory, and sandbox environments to fit their specific technical stack.

Native Sandbox Execution

The updated SDK supports native sandbox execution, allowing agents to run in controlled computer environments with necessary files, tools, and dependencies. This removes the requirement for developers to manually piece together an execution layer.

Sandbox Provider Support

Developers can utilize built-in support for several providers, including:

  • Blaxel
  • Cloudflare
  • Daytona
  • E2B
  • Modal
  • Runloop
  • Vercel

Workspace Portability via Manifests

To ensure environments are portable across different providers, the SDK introduces a Manifest abstraction. This allows developers to define the agent's workspace by mounting local files, defining output directories, and importing data from cloud storage providers such as AWS S3, Google Cloud Storage, Azure Blob Storage, and Cloudflare R2.

This Manifest system provides the model with a predictable workspace, ensuring it knows where to find inputs and write outputs across long-running tasks.

Security, Durability, and Scalability

By separating the harness from the compute layer, the Agents SDK addresses critical production concerns:

  • Security: Separating harness and compute keeps credentials out of environments where model-generated code executes, mitigating risks associated with prompt-injection and exfiltration.
  • Durability: The SDK supports built-in snapshotting and rehydration. If a sandbox container fails or expires, the agent's state can be restored in a fresh container to continue from the last checkpoint.
  • Scalability: Agent runs can parallelize work across multiple containers, route subagents to isolated environments, or invoke sandboxes only as needed.

Availability and Roadmap

These capabilities are generally available to all customers via the API using standard token and tool-use pricing.

Currently, the new harness and sandbox capabilities are available in Python, with TypeScript support planned for a future release. Future updates will include the added support for "code mode" and subagents for both languages.

Sources