OpenAI updates Agents SDK

OpenAI updated its Agents software development kit on April 15, adding native sandbox execution and a new runtime for agents that work across files and tools. (openai.com) An agent is software that can call tools and take multiple steps instead of answering in one shot. OpenAI’s updated kit gives those agents a controlled workspace where they can inspect files, run commands, edit code, and keep working on longer tasks. (openai.com) The company said the new runtime is a “model-native harness,” meaning the surrounding control system is tuned for OpenAI models rather than bolted on afterward. OpenAI also said the release is available through its application programming interface at standard pricing. (openai.com) (techcrunch.com) The security piece is the sandbox. OpenAI said developers can give an agent a limited workspace and explicit tools so it touches only the files, code, and commands needed for a task instead of running freely on a company system. (openai.com) (techcrunch.com) The software also adds built-in tracing, which is a recording layer for watching what an agent did step by step. OpenAI’s documentation says developers can use that tracing to visualize, debug, monitor, and evaluate agent workflows. (openai.github.io) That matters for handoffs, where one agent passes work to another agent with a narrower job. OpenAI’s documentation lists handoffs, sessions, guardrails, and human-in-the-loop controls as core parts of the kit for multi-step workflows. (openai.github.io) OpenAI framed the update as a response to production problems that show up after a prototype works. In its launch post, the company said teams often face tradeoffs between flexible model-agnostic frameworks, provider software development kits with limited visibility, and managed agent services that restrict where agents run or how they access sensitive data. (openai.com) TechCrunch reported that the first release of the new harness and sandbox features is in Python, with TypeScript support planned later. Karan Sharma, who works on OpenAI’s product team, told the publication the goal is to let companies build “long-horizon agents” on their own infrastructure. (techcrunch.com) OpenAI’s public documentation says the kit sits above the Responses application programming interface and is meant for cases where developers want the runtime to manage turns, tool execution, guardrails, handoffs, or sessions. For shorter workflows where developers want to control the loop themselves, OpenAI says to use the Responses interface directly. (openai.github.io) The release pushes OpenAI’s agent tools closer to the work enterprises usually care about: file handling, approvals, monitoring, and resumable execution inside bounded environments. OpenAI said it will keep expanding the kit over time. (openai.com) (techcrunch.com)