OpenAI upgrades its open-source Agents SDK with enterprise-grade sandboxing, a long-horizon harness, subagents, code mode, and support for 100+ LLMs.
OpenAI's Agents SDK — the production successor to the experimental Swarm project — receives its most significant update yet. New capabilities address the two hardest enterprise concerns: safety during autonomous operation, and support for complex multi-step work.
Two capabilities ship now in Python; two more are on the roadmap for both Python and TypeScript.
Every Agents SDK workflow is built from four composable building blocks.
Three APIs matter for agent development. The Responses API unifies the best of the others and is the recommended foundation going forward. The Assistants API is being deprecated in mid-2026.
pip install openai-agents. Set OPENAI_API_KEY in your environment. For TypeScript: npm install @openai/agents. Available to all OpenAI API customers at standard pricing.@function_tool or built-in tools like web search and code execution.A lightweight, production-ready open-source framework for building multi-agent workflows in Python and TypeScript. Evolved from the experimental Swarm project. Supports OpenAI APIs and 100+ third-party LLMs.
Sandboxing isolates the agent inside a controlled computer environment where it can only access designated files and code. This protects the broader host system from unpredictable agent behaviour during autonomous operation.
New orchestration infrastructure for building long-horizon agents — complex, multi-step tasks. Developers bring their own compute and storage; the harness provides the coordination layer. Launching in Python first.
(1) Agents — LLM workers with instructions and tools; (2) Handoffs — agent-to-agent task delegation; (3) Guardrails — I/O validation for safety and policy compliance. Plus built-in Tracing.
The Agents SDK works with any model that exposes a Chat Completions-compatible API endpoint — enabling 100+ non-OpenAI LLMs from third-party and open-source providers.
It combines the language capability of the Chat Completions API with the tool-use functionality of the Assistants API in one modern API. The Assistants API will be deprecated in mid-2026.
A primary agent delegates a subtask to another specialised agent. The receiving agent has its own instructions and tools. Agents can also be exposed as tools ('Agents as Tools'), letting an orchestrator call them like any other function.
Subagents allow an orchestrator to spawn specialised subordinate agents for parallel, modular task decomposition. They are forthcoming — being added to both Python and TypeScript versions of the SDK.
Swarm was experimental and not production-ready. The Agents SDK adds formal primitives (handoffs, guardrails, tracing), enterprise safety (sandboxing), long-horizon support (harness), and provider-agnostic model compatibility.
The Agents SDK itself is open-source and free. It is available to all OpenAI API customers. API calls use standard OpenAI pricing.
The core Agents SDK is available in both Python and TypeScript. The new sandboxing and harness capabilities launch in Python first; TypeScript support is planned. Subagents and code mode will come to both.