openai/openai-agents-python

A lightweight, powerful framework for multi-agent workflows

What it solves

It provides a lightweight framework for building multi-agent workflows, allowing developers to create complex AI systems where multiple specialized agents can collaborate, hand off tasks to one another, and interact with external tools or sandboxed environments.

How it works

The SDK uses a provider-agnostic approach that supports OpenAI APIs and over 100 other LLMs. It organizes functionality around several core concepts:

  • Agents: LLMs configured with specific instructions, tools, and guardrails.
  • Handoffs: The ability for one agent to delegate a task to another agent.
  • Sandbox Agents: Specialized agents that can operate within a container to run commands, inspect files, and maintain state over long periods.
  • Tools: Integration with functions, MCP, and hosted tools to allow agents to take actions.
  • Realtime Agents: Support for low-latency voice and multimodal experiences via WebSockets using gpt-realtime-2.1.
  • Infrastructure: Built-in session management for conversation history, tracing for debugging, and human-in-the-loop mechanisms for validation.

Who it’s for

Developers building AI agents and multi-agent systems who need a standardized way to manage handoffs, tool use, and persistent workspaces across different LLM providers.

Highlights

  • Provider Agnostic: Works with OpenAI and 100+ other LLMs.
  • Sandbox Integration: Built-in support for agents to perform work in isolated containers.
  • Realtime Voice: Native support for low-latency voice agents.
  • Multi-Agent Orchestration: Native handoff mechanisms to delegate tasks between agents.
  • Developer Tooling: Includes built-in tracing and session history management.