OpenAI is releasing an updated version of its Agents SDK. The update is aimed at developers who want to build AI agents capable of performing complex tasks autonomously.
The new version emphasizes practical usability and introduces, among other things, an integrated sandbox environment and a more comprehensive foundation for long-running and multi-step processes. This move is part of a broader trend in which companies like OpenAI and Anthropic are competing to provide businesses with tools for so-called agentic AI.
With the new SDK, agents can not only generate text but also work with files, execute code, and invoke system commands. This shifts the role of AI from assistant to an executive system that handles tasks within a controlled environment. According to OpenAI, developers have often lacked consistent infrastructure to reliably deploy these types of applications in production until now.
A key component is the so-called harness, the layer that organizes interactions among the model, tools, and data. This term refers to all the components surrounding the model that are necessary for an agent to function, as TechCrunch also explains. The new version is better tailored to how advanced models operate, which, according to OpenAI, should contribute to more stable performance on more complex tasks.
In addition, OpenAI is introducing standard support for sandbox execution. Agents operate in a protected environment where they safely execute code and work with files without direct access to sensitive systems. This is relevant because agents sometimes exhibit unpredictable behavior. Within the sandbox, they work in isolation in a specific working environment with only authorized data and tools, according to TechCrunch.
A middle ground between open frameworks and closed systems
The update addresses a well-known tension. Generic frameworks offer flexibility but do not fully leverage the capabilities of advanced models. At the same time, tools from model providers are often less transparent, while fully managed solutions impose limitations. OpenAI positions the SDK as a middle ground that combines control and ease of use.
The infrastructure is designed for so-called long-horizon tasks, in which agents go through multiple steps and combine different tools. Such processes require a system to function consistently over an extended period.
The separation between the agent’s logic and the computing environment also affects security and scalability. Sensitive data remains outside the execution environment, and tasks can be resumed if an environment fails. This enables agents to run for extended periods and distribute work across multiple environments.
According to companies that tested the SDK, the update enables automating processes that were previously too complex or error-prone. This involves not only extracting information but also correctly interpreting context within large datasets.
The new functionality is available via the API and is offered under the existing pricing model. For now, the SDK is only available in Python, with TypeScript support in development. OpenAI is working on further enhancements, including collaboration between multiple agents and deeper integrations with external systems.