OpenAI updates Agents SDK

An AI agent is software that can take a task, use tools, and keep working across multiple steps instead of answering in one shot. On April 15, OpenAI updated its Agents software development kit so those agents can run code and handle files inside controlled workspaces. (openai.com) The update adds two core pieces: a model-native harness, which is the scaffolding around the model, and native sandbox execution, which isolates the agent while it works. OpenAI said the harness lets agents operate across files and tools on a computer, while the sandbox lets them run that work more safely. (openai.com) OpenAI said the new harness and sandbox features launched first in Python on April 15, with TypeScript support planned for a later release. The company said the capabilities are generally available through the application programming interface, with standard pricing based on tokens and tool use. (helpnetsecurity.com) The harness is meant to solve a specific engineering problem: agents that inspect documents, run shell commands, edit code, and keep state over long jobs need more than a model call. OpenAI said existing options force tradeoffs between flexibility, visibility into the runtime, and control over where sensitive data is processed. (openai.com) The sandbox is the safety layer in that stack. OpenAI and TechCrunch described it as a siloed environment where an agent gets access only to the files, tools, and dependencies needed for a task, instead of operating directly on a broader system. (openai.com) (techcrunch.com) OpenAI said the updated harness includes configurable memory, sandbox-aware orchestration, filesystem tools similar to Codex, and integrations for common agent primitives. The company listed tool use via Model Context Protocol, skills, AGENTS.md instructions, shell execution, and apply_patch-style file edits among the supported patterns. (helpnetsecurity.com) The company also said developers can bring their own sandbox or use built-in support for providers including Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel. Karan Sharma, an OpenAI product manager, told TechCrunch the goal was to make the kit work with different sandbox providers so teams can build long-running agents on their own infrastructure. (helpnetsecurity.com) (techcrunch.com) OpenAI used customer examples to frame the release as a production tool, not a lab demo. The company published comments from Oscar Health and LexisNexis saying the updated kit helped with clinical-record workflows and legal drafting tasks that require secure, isolated processing and many steps. (openai.com) (helpnetsecurity.com) OpenAI said more features, including code mode and subagents, are in development for Python and TypeScript. For developers trying to move from chatbots to software that can actually inspect files, run commands, and finish longer jobs, the April 15 release is OpenAI’s latest attempt to provide the runtime around the model, not just the model itself. (helpnetsecurity.com) (openai.com)