OpenAI Upgrades Agent Tools for Developers

OpenAI is rolling out a significant update to its Agents SDK, aiming to empower developers to build more sophisticated AI agents capable of complex tasks. The enhanced SDK provides a standardized infrastructure designed to work seamlessly with OpenAI's models, enabling agents to inspect files, execute commands, and manage long-horizon projects within controlled environments.

A More Capable Agent Harness

The core of the update lies in a more capable harness, which now includes configurable memory, sandbox-aware orchestration, and file system tools. This allows agents to interact with documents and systems more effectively.

This new harness integrates with emerging primitives in agent systems, such as tool use via MCP and progressive disclosure via skills. It also supports custom instructions and code execution through shell commands and file edits, reducing the need for developers to build this core infrastructure themselves.

By aligning agent execution with how frontier models operate best, the SDK promises improved reliability and performance, especially for intricate, long-running tasks. OpenAI emphasizes flexibility, allowing developers to adapt the harness to their specific stacks.

Native Sandbox Execution for Safety

A key addition is native sandbox execution, providing agents with secure, controlled environments to perform tasks. This is crucial for agents that need to read/write files, install dependencies, or run code.

The SDK offers out-of-the-box support for this execution layer, eliminating the need for developers to assemble it manually. It integrates with various sandbox providers, including Blaxel, Cloudflare, and Vercel, and introduces a Manifest abstraction for portable workspace descriptions. This ensures a consistent environment from local development to production deployment.

This feature is critical for applications requiring secure agent operations, potentially boosting security through features discussed in Agent Sandboxing Boosts Security, and aligns with advancements in platforms like Cloudflare Sandboxes Go General.

Separating Harness from Compute

OpenAI highlights the importance of separating the agent harness from compute for security and durability. This architecture helps prevent credentials from being exposed in execution environments.

Externalizing the agent's state enables durable execution; losing a sandbox container does not mean losing the entire run. Built-in snapshotting and rehydration allow agents to resume tasks in new environments from their last checkpoint.

This separation also enhances scalability, allowing agent runs to utilize multiple sandboxes, route subagents to isolated environments, and parallelize work for faster results.

Availability and Future Plans

The new capabilities are available now via the API with standard API pricing. Initially launching in Python, TypeScript support is planned for a future release, alongside further enhancements like code mode and subagents.

OpenAI intends to foster the broader agent ecosystem by supporting more sandbox providers and integrations, enabling developers to plug the SDK into their existing toolchains.