OpenAI updates Agents SDK

OpenAI has added built-in sandboxing and a testing harness to its Agents Software Development Kit, aiming the update at companies deploying autonomous software. (openai.com) The company announced the changes on April 15, 2026, and said the new release lets developers run agent actions inside isolated environments and evaluate behavior before production rollout. (openai.com) An agent is software that can plan steps, use tools, and keep state across a task; OpenAI’s documentation says the Agents Software Development Kit is designed for applications that manage orchestration, tool execution, approvals, and memory in code. (developers.openai.com) Sandboxing is the software equivalent of giving a contractor a locked workroom instead of keys to the whole building. OpenAI said its new native sandbox execution is meant to let agents work across files and tools while keeping that work contained. (openai.com) The new harness is a test rig for agents before they go live. OpenAI said the model-native harness helps developers measure whether an agent stays inside expected behavior on tasks drawn from the same distribution as real production work. (openai.com; developers.openai.com) The release extends a product line OpenAI began rolling out in March 2025, when it introduced the Responses Application Programming Interface, built-in tools such as web and file search, and the first Agents Software Development Kit. (openai.com) Since then, OpenAI has been pushing developers toward more formal evaluation. Its agent evaluation guide now centers on traces, graders, datasets, and evaluation runs, and a 2025 cookbook example described offline and online metrics for getting agents into production. (developers.openai.com; developers.openai.com) The enterprise pitch is straightforward: companies want agents that can touch files, run commands, and use tools, but they also want logs, limits, and repeatable tests before those systems reach customer data or internal systems. OpenAI’s current Agents overview says sandbox agents in Python can use containers with files, commands, packages, ports, snapshots, and memory. (developers.openai.com) OpenAI has also been building adjacent products for the same audience. Its agent platform page now offers a code-first Agents Software Development Kit alongside a visual workflow product called Agent Builder, both tied to the Responses Application Programming Interface. (openai.com) The update does not remove the core tradeoff around agent software: the more useful the system becomes, the more carefully companies have to constrain where it can act. This release gives those controls a more official place inside OpenAI’s own stack. (openai.com; developers.openai.com)