OpenAI's Toolkit Update Aims to Tame Unpredictable AI Agents

OpenAI is refining the tools it provides to businesses building AI agents. The company has released a significant update to its Agents SDK, a toolkit for creating automated assistants powered by its models, with a clear focus on safety and control for enterprise environments.

The update introduces a sandboxing feature, allowing these agents to operate in isolated computer workspaces. This addresses a core concern for developers: an agent acting unpredictably could pose a risk if given unrestricted system access. Sandboxing confines an agent's actions to a specific area, letting it read files or execute code for a task without threatening the broader system's stability.

Accompanying this is a new 'in-distribution harness' for what the industry calls frontier models—the most advanced and capable AI systems available. This harness provides the necessary framework to deploy and test these powerful models within a workspace, connecting them safely to approved files and tools.

"We're taking our existing agents SDK and making it compatible with all of these sandbox providers," explained Karan Sharma of OpenAI's product team. The goal, he noted, is to let users construct "long-horizon" agents—capable of complex, multi-step projects—using OpenAI's framework and their own infrastructure.

The new harness and sandbox capabilities are available first in Python, with TypeScript support expected later. OpenAI plans to add more features, such as code mode and subagents, to both languages. All customers can access these updates through the standard API with existing pricing.

Source: TechCrunch