OpenAI updates Agents SDK
OpenAI released an update to its Agents SDK that adds 'harness' and 'sandbox' capabilities to help enterprises build safer, bounded agents, with the first release in Python and TypeScript support promised later. The SDK focuses on controlled execution, testing and isolation for agentic workflows used in production settings. (techcrunch.com)
OpenAI has expanded its Agents software development kit with a new “harness” and a native sandbox, pushing more of the safety and execution layer into the toolkit itself. (openai.com) An agent is software that can plan a task, call tools, and keep working across multiple steps instead of answering with one chat response. OpenAI says its Agents software development kit now gives developers a model-native harness for instructions, approvals, tracing, handoffs, and resumable work. (developers.openai.com, openai.com) The sandbox is the separate computer environment where that work runs. OpenAI says sandbox agents can manipulate files, run commands, install packages, expose services, and save state for later, while staying isolated from the main application. (developers.openai.com) OpenAI released the sandbox capability first in the Python Agents software development kit on April 15, 2026. The company said TypeScript support is planned next, while its developer site already lists Agents software development kit resources for both Python and TypeScript. (openai.com, developers.openai.com) The update targets companies that want agents to do work on computers without giving them open-ended access to production systems. OpenAI said developers can run the harness and the compute layer on the same machine or split them apart, which lets teams keep tighter boundaries around what an agent can touch. (openai.com, developers.openai.com) OpenAI has been moving its developer stack toward agent workflows for more than a year. In March 2025, the company introduced the Responses application programming interface and new built-in tools as the foundation for agentic applications. (openai.com) The Agents software development kit itself grew out of earlier work called Swarm. OpenAI’s Python documentation describes the kit as a production-ready upgrade of that earlier experiment, with support for multi-agent handoffs and full traces of what happened. (openai.github.io, developers.openai.com) OpenAI framed the new release as infrastructure for long-running agents that work across files and tools on a computer, not just chat windows. TechCrunch reported the company is aiming the changes at enterprise deployments where controlled execution and isolation are required before agents can be used in production settings. (openai.com, techcrunch.com) The immediate next step is practical: Python developers can start using sandboxes now, and OpenAI is promising the same bounded setup for TypeScript later. That keeps the focus on the same problem the update is trying to solve — letting agents do more work while keeping that work contained. (developers.openai.com, openai.com)
