OpenAI updates Agents SDK for building safer, smarter AI agents
What's the story
OpenAI has released a major update to its agent software development kit (SDK), aimed at improving the safety and reliability of AI tools for enterprises. The latest version comes with sandboxing capabilities, which keep agents within safe zones to prevent them from going off-script or causing unexpected issues. It also features an in-distribution harness that enables agents to handle files and tools more precisely.
Features explained
In-distribution harness for frontier models
The new SDK update offers developers an in-distribution harness for frontier models, allowing agents to work with files and approved tools within a workspace. This feature is particularly useful in agent development as it provides both deployment and testing capabilities for agents running on advanced, general-purpose models. The idea behind this launch is to make the existing agents SDK compatible with all sandbox providers.
Future prospects
Compatibility with all sandbox providers
Karan Sharma from OpenAI's product team said, "This launch, at its core, is about taking our existing agents SDK and making it so it's compatible with all of these sandbox providers." He hopes that this, along with the new harness capabilities, will let users "to go build these long-horizon agents using our harness and with whatever infrastructure they have." Long-horizon tasks are usually more complex and multi-step.
Future expansion
Future plans for the SDK
OpenAI has promised to keep expanding the Agents SDK over time. However, the new harness and sandbox capabilities are initially being launched only in Python, with TypeScript support planned for a later release. The company is also working on bringing more agent capabilities such as code mode and subagents to both Python and TypeScript. These new features can be accessed by all customers via the API at standard pricing.