OpenAI Agents SDK (Python)

Use it when

• You are hand-rolling agent loops, tool dispatch, handoff logic, and conversation history for a multi-agent Python application
• You need voice, realtime, or sandboxed coding agents with built-in guardrails, human-in-the-loop, and run tracing without building that plumbing yourself

What it solves

Not the fit when

• Python 3.10+ only; JavaScript and TypeScript users need the separate Agents SDK JS/TS project
• The retained quickstart examples require OPENAI_API_KEY even though the SDK also supports other LLM providers
• UnixLocalSandboxClient is supported on macOS and Linux; Windows requires DockerSandboxClient with the docker extra or a hosted sandbox client
• Agent tools and sandbox agents can take actions, run commands, and modify files; add guardrails and human approval appropriate to the actions you expose
• local model serving and inference
• browser automation
• vector retrieval and RAG indexing
• provider cost routing on its own

Install

pip install openai-agents  # Python 3.10+; extras: 'openai-agents[voice]' for voice, '[redis]' for Redis sessions, '[docker]' for Docker sandbox

Invoke

Set OPENAI_API_KEY, then: from agents import Agent, Runner; agent = Agent(name='Assistant', instructions='...'); Runner.run_sync(agent, 'prompt').final_output; use SandboxAgent for workspace tasks, RealtimeAgent for voice over WebSocket, VoicePipeline for STT-agent-TTS

Alternatives

No reviewed alternatives recorded yet.