OpenAI has released a significant update to its Agents SDK, shifting its focus from prototyping to production. The new version introduces two core architectural features aimed at engineers building systems with autonomous agents.

The first is sandbox integration. This allows an agent to execute within a controlled, isolated environment—a critical design decision for managing risk. Instead of granting an agent broad system access, it can be confined to a specific workspace with defined permissions for files and code execution. This containment addresses the operational unpredictability inherent in LLM-driven automation.

Second is the introduction of an 'in-distribution harness' for frontier models like GPT-4. In practical terms, this harness is the framework that surrounds the core model, managing its interactions with tools, files, and external systems within a workspace. It provides a standardized interface for both deploying and testing these complex agents. According to OpenAI's Karan Sharma, the goal is to let users build 'long-horizon' agents—capable of multi-step, intricate tasks—using this harness atop their existing infrastructure.

For engineering teams, the trade-off is clear: adopting this managed harness may accelerate development but couples the agent's orchestration layer to OpenAI's tooling. The update is available now in Python via the standard API, with TypeScript support and advanced features like subagents planned for later releases. Pricing remains unchanged.

Source: TechCrunch