// the find
openai/openai-agents-python
A lightweight, powerful framework for multi-agent workflows
OpenAI's official Python SDK for building multi-agent workflows — think orchestrators that delegate to specialized sub-agents, with built-in tracing, guardrails, and MCP support. It's provider-agnostic despite the name, routing through LiteLLM to hit 100+ models. Aimed at developers who want structure around agent loops without writing the plumbing from scratch.
First-party maintenance means it stays in sync with OpenAI API changes the day they ship, which community wrappers never quite manage. The tracing layer is genuinely useful — run-level visibility with a UI, not just logging to stdout. MCP integration is built in rather than bolted on, so tool servers are a first-class primitive. Sandbox agents (v0.14+) give agents a real filesystem and shell rather than simulated tool calls, which closes the gap between demo and actually-useful for code tasks.
Provider-agnostic is the claim, but the abstractions still leak OpenAI-shaped assumptions — context window handling, tool call formats, and streaming behavior all have OpenAI-flavored edge cases that surface when you swap to Anthropic or Gemini. The handoff model assumes agents know at configuration time who they can delegate to; dynamic agent graphs are awkward. Session persistence is SQLite or Redis only out of the box, so anyone on Postgres is writing their own adapter from day one. The sandbox agents are new enough that the non-local clients (E2B, Daytona, Cloudflare, etc.) are thin wrappers with minimal real-world validation visible in the issue tracker.