OpenAI Updates Agents SDK for Safer Enterprise Agents
OpenAI has refreshed its Agents software development kit. The changes aim to assist companies in creating agents that rely on OpenAI's models. Agentic AI has become a key focus in the tech sector. Firms such as OpenAI and Anthropic compete to supply enterprises with resources for these automated assistants.
Sandboxing for Controlled Agent Operations
A major addition is sandboxing. This feature lets agents function in isolated computer settings. Such isolation matters because agents can act in unexpected ways without oversight. Risks arise from their autonomous behavior.
In a sandbox, agents stay confined to a specific workspace. They can access files and code only for assigned tasks. The rest of the system remains protected from potential issues.
In-Distribution Harness for Frontier Models
The updated SDK also includes an in-distribution harness for frontier models. Developers can use it to handle files and permitted tools inside a workspace, according to OpenAI. In agent development, a harness covers all parts of an agent except the underlying model. An in-distribution version supports deployment and testing on frontier models. These represent the top tier of advanced, versatile models available.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Karan Sharma from OpenAI's product team explained the focus. "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 told TechCrunch.
He added that combined with the harness, users can create long-horizon agents. They can use OpenAI's harness alongside their own setups. Long-horizon tasks involve complex, multi-step processes.
Future Expansions and Availability
OpenAI plans ongoing improvements to the Agents SDK. The sandbox and harness features debut in Python. TypeScript support will follow later. Additional functions, such as code mode and subagents, are in development for both languages.
All customers access these via the API. Pricing follows standard rates.
OpenAI, founded in 2015 as a nonprofit research lab, shifted to a capped-profit model in 2019 to attract investment. It gained massive attention with the 2022 launch of ChatGPT, built on GPT-3.5 and later GPT-4 models. The company leads in generative AI, powering tools used by millions. Its frontier models, like GPT-4o, handle diverse tasks from text to vision. Agentic systems build on this by enabling AI to take actions autonomously, such as booking travel or coding apps. Enterprises seek these for efficiency but demand safeguards against errors or misuse. OpenAI's SDK update addresses that need directly, fitting into broader efforts like its o1 reasoning models released in 2024.
Anthropic, a rival founded by ex-OpenAI staff in 2021, offers Claude models with similar agent tools. The race underscores growing enterprise demand for reliable AI agents amid regulatory scrutiny on safety.