OpenAI Agents SDK

The OpenAI Agents SDK is the official framework from OpenAI for building multi-agent systems in Python. Despite the name, it is provider-agnostic: it works with OpenAI APIs plus over 100 other LLMs, so you are not locked into a single provider. The SDK revolves around nine core concepts. Agents are LLMs configured with instructions, tools, guardrails, and handoffs. Tools can be Python functions, MCP servers, or hosted tools. Guardrails provide configurable safety checks for input and output validation. Handoffs let agents delegate tasks to other agents. Human-in-the-loop mechanisms let you involve people at any step. Sessions manage conversation history automatically across runs. Tracing tracks every step for viewing, debugging, and optimization. Realtime agents support voice workflows with gpt-realtime-1.5. The newest addition is Sandbox Agents (v0.14.0+). A sandbox agent operates in a containerized environment with its own filesystem, shell access, and workspace. You define a Manifest specifying files, directories, Git repos, environment variables, and mounts. The agent can then inspect files, run commands, apply patches, and carry workspace state across longer tasks. Sandbox providers include local Unix, Docker, and hosted options (Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, Vercel). Sandbox Memory lets future runs learn from prior runs with read-only and generate-only modes, live updates when stale memory is discovered, and multi-turn grouping. Workspace mounts support local files, S3, Cloudflare R2, GCS, and Azure Blob Storage. Installation is straightforward: `pip install openai-agents`. Optional extras include voice support and Redis sessions. The SDK requires Python 3.10+ and is MIT licensed with 25,400 GitHub stars and 278 contributors.