The next evolution of the Agents SDK

Source: OpenAI Blog

Introduction

We’re introducing new capabilities to the Agents SDK that give developers standardized infrastructure that is easy to get started with and is built correctly for OpenAI models: a model‑native harness that lets agents work across files and tools on a computer, plus native sandbox execution for running that work safely.

Developers need more than the best models to build useful agents—they need systems that support how agents inspect files, run commands, write code, and keep working across many steps.

Limitations of Existing Systems

The systems that exist today come with trade‑offs as teams move from prototypes to production:

• Model‑agnostic frameworks – flexible but do not fully utilize frontier model capabilities.
• Model‑provider SDKs – closer to the model but often lack enough visibility into the harness.
• Managed agent APIs – simplify deployment but constrain where agents run and how they access sensitive data.

New SDK Capabilities

With today’s release, the Agents SDK harness becomes more capable for agents that work with documents, files, and systems. It now includes:

• Configurable memory
• Sandbox‑aware orchestration
• Codex‑like filesystem tools
• Standardized integrations with primitives common in frontier agent systems

The harness aligns execution with the way frontier models perform best, keeping agents closer to the model’s natural operating pattern and improving reliability and performance on complex, long‑running, or coordinated tasks.

Flexibility for Diverse Products

We designed the Agents SDK to support diverse product requirements. Developers get a turnkey yet flexible harness that can be adapted to their own stack, including:

• Custom tool use
• Adjustable memory settings
• Choice of sandbox environment

Native Sandbox Execution

The updated SDK supports sandbox execution natively, allowing agents to run in controlled computer environments with the files, tools, and dependencies they need for a task.

• Built‑in sandbox providers: Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, Vercel
• Custom sandboxes: Developers can bring their own

Manifest Abstraction

To make environments portable across providers, the SDK introduces a Manifest abstraction for describing the agent’s workspace. With a manifest you can:

• Mount local files
• Define output directories
• Import data from storage providers (AWS S3, Google Cloud Storage, Azure Blob Storage, Cloudflare R2)

This gives a consistent way to shape the agent’s environment from prototype to production and provides the model with a predictable workspace.

Durability and Scalability

• Prompt‑injection and exfiltration protection: Separating harness and compute keeps credentials out of environments where model‑generated code executes.
• Durable execution: Agent state is externalized; built‑in snapshotting and rehydration allow restoration in a fresh container if the original environment fails or expires.
• Scalable execution: Agents can use one or many sandboxes, invoke sandboxes on demand, route subagents to isolated environments, and parallelize work across containers for faster execution.

Availability and Pricing

These new Agents SDK capabilities are generally available to all customers via the API and use standard API pricing, based on tokens and tool use.

Roadmap

• Current launch: Harness and sandbox capabilities are available in Python.
• Upcoming: TypeScript support, code mode, and subagents for both Python and TypeScript.
• Future plans: Expand sandbox provider support, add more integrations, and provide additional ways for developers to plug the SDK into existing tools and systems.

Customer Feedback

“With today’s release, the Agents SDK harness becomes more capable for agents that work with documents, files, and systems. It now has configurable memory, sandbox‑aware orchestration, Codex‑like filesystem tools, and standardized integrations with primitives that are becoming common in frontier agent systems.”

“The harness also helps developers unlock more of a frontier model’s capability by aligning execution with the way those models perform best, improving reliability and performance on complex tasks.”

These testimonials illustrate the practical benefits observed by early adopters.