OpenAI Agents SDK

OpenAI Agents SDK is an MIT-licensed, Python-first framework for building applications in which language models use tools, delegate work and continue through a managed agent loop. Its central abstractions are agents, tools, handoffs and guardrails, with the runtime handling turns and tool execution while application code retains control over orchestration. OpenAI models use the Responses API by default, and the model layer also documents routes to other providers. The package succeeds OpenAI's experimental Swarm project and is intended for developers who want more structure than calling a model API directly without adopting a large workflow system.

Top features

• Define agents with instructions, tools, typed outputs, handoffs and input or output guardrails. Python functions can become schema-backed tools, while hosted tools and agents-as-tools cover broader delegation patterns.
• Build multi-agent systems either by handing control to a specialist or by keeping a manager agent in charge while specialists run as tools.
• Connect agents to Model Context Protocol servers, with documented transports, tool filtering, prompt access, caching and trace integration.
• Keep conversational state through sessions and choose how prior turns are carried into a run. Runs can pause for human approval and resume after a decision.
• Inspect model generations, tool calls, handoffs, guardrails and custom events through built-in tracing. Hooks expose lifecycle events for additional logging or application behavior.
• Create realtime voice agents or speech-to-text, agent and text-to-speech pipelines. Sandbox agents add isolated, resumable workspaces for tasks that need files or shell capabilities.

Use cases

• Route support, intake or operations requests to specialized agents while preserving a clear final-response owner.
• Build research and document workflows that call internal functions, hosted tools or MCP services and retain context between turns.
• Add validation and approval checkpoints around sensitive tool calls or structured outputs.
• Prototype and operate low-latency voice assistants, or run coding and review tasks inside an isolated workspace.
• Trace multi-step behavior during development, evaluation and production monitoring without designing a separate event model first.