openai/openai-agents-python

What it solves

This SDK provides a lightweight framework for building complex multi-agent workflows. It simplifies the orchestration of multiple LLMs, allowing them to delegate tasks to one another, manage conversation history, and interact with external tools and environments.

How it works

The framework is provider-agnostic and supports OpenAI's APIs as well as over 100 other LLMs. It organizes AI behavior through several specialized agent types:

• Text Agents: Standard LLM-based assistants for general text workflows.
• Sandbox Agents: Agents that operate within a containerized workspace to run commands, inspect files, and maintain state over long-term tasks.
• Realtime Agents: Low-latency agents designed for voice and multimodal experiences using WebSockets.
• Voice Agents: Specialized pipelines that integrate speech-to-text, agent logic, and text-to-speech.

It also includes built-in mechanisms for handoffs (delegating to other agents), guardrails for safety validation, and "Human in the loop" support for manual intervention.

Who it’s for

Developers building AI agents that require sophisticated orchestration, voice integration, or the agent's ability to interact with a local or hosted file system/terminal (sandboxes).

Highlights

• Multi-Agent Orchestration: Built-in support for agent-to-agent handoffs and treating agents as tools.
• Versatile Modalities: Native support for text, real-time voice, and multimodal interactions.
• Sandbox Integration: Ability to run agents in secure containers to perform technical tasks like applying patches or running code.
• Developer Tooling: Integrated tracing for debugging and session management for automatic history tracking.