OpenAI open-sources Agents SDK
- OpenAI has open-sourced its Agents SDK, a code-first framework for building AI agents with tools, handoffs, guardrails, tracing, and human review. - The SDK’s core pitch is small and opinionated: a few primitives — agents, tools, handoffs, sessions, and traces — instead of a giant workflow stack. - That matters because teams want production agents, not demos — and OpenAI is trying to make its orchestration layer a default starting point.
OpenAI has pushed its Agents SDK into the open as a real developer product, not just an internal pattern with docs wrapped around it. The point is simple: most “agents” break down at the exact moment they need to do multi-step work, call tools, hand tasks to another specialist, and leave behind a debuggable trail. OpenAI is trying to turn that messy middle into a standard kit. The result is a lightweight framework for Python and JavaScript that handles orchestration, guardrails, tracing, and handoffs in one place. ### What is the SDK actually for? It’s for the part after the model call. A normal API client gets you a response. An agent app has to decide what tool to call, whether to ask a human, whether another agent should take over, how to keep state across turns, and how to inspect what happened when something goes wrong. The Agents SDK is meant for teams that want to own that orchestration in code instead of using a hosted workflow editor. ### What are the core building blocks? OpenAI keeps the vocabulary deliberately small. An agent is a model plus instructions and tools. Handoffs let one agent delegate work to another. Guardrails validate inputs and outputs. Sessions keep conversation state. Tracing records the run so developers can inspect and debug it later. There’s also human review support for workflows that should pause before risky actions continue. ### Why does “open-source” matter here? Because orchestration code is where teams get locked in. If the framework is public, developers can inspect the harness, extend it, and run it inside their own stack. OpenAI’s GitHub repos show the SDKs are MIT-licensed and active — the Python repo has roughly 26k stars, while the JavaScript repo is near 3k. That doesn’t prove production adoption by itself, developers can build on. ### Is this only for OpenAI models? Not entirely. The Python SDK explicitly says it is provider-agnostic and supports the OpenAI Responses and Chat Completions APIs plus 100+ other LLMs. That’s an important detail. OpenAI wants the framework to feel broader than a thin wrapper around its own API, even while the docs clearly steer users toward OpenAI-native tools and runtime patterns. Basically, it’s best with OpenAI’s stack. ### What changed more recently? The bigger April 15, 2026 update pushed the
