OpenAI Agents SDK

The OpenAI Agents SDK is a lightweight but opinionated framework for building multi-agent workflows in Python. It structures applications around a small set of primitives — Agents (LLMs configured with instructions, tools, and guardrails), Handoffs (delegating a task to another agent), Sessions (automatic conversation history), and Tracing (built-in observability for every agent run) — and is provider-agnostic, working with OpenAI’s Responses and Chat Completions APIs as well as 100+ other LLMs via LiteLLM and other adapters.

Beyond the core agent loop, it ships batteries-included support for function tools, hosted tools, MCP (Model Context Protocol) servers, sandboxed “sandbox agents” for long-running container-backed work, realtime voice agents, and human-in-the-loop approval flows. As the official successor to OpenAI’s earlier Swarm experiment, it’s become one of the most widely adopted agent frameworks, backing production multi-agent systems that need structured orchestration rather than ad-hoc prompt chaining.

What You Get

• An Agent primitive combining instructions, model settings, tools, and guardrails into a reusable unit
• Handoffs and agents-as-tools for delegating subtasks between agents in a multi-agent workflow
• Built-in Sessions for automatic conversation history management across runs
• Native tracing that records every agent run for debugging and workflow optimization
• Support for function tools, hosted tools, and MCP servers so agents can take real actions
• Sandbox Agents preconfigured to operate inside containers (Docker, E2B, Modal, Daytona, Cloudflare, etc.) for long-horizon tasks
• Provider-agnostic model layer supporting OpenAI Responses/Chat Completions plus 100+ other LLMs via LiteLLM

Common Use Cases

• Building a customer-support system where a triage agent hands off to specialized agents (billing, technical, refunds)
• Orchestrating a research or coding agent that uses tools and sandboxed execution over a long-running task
• Adding conversational memory and tracing/observability to a production LLM application
• Building realtime voice agents or human-in-the-loop approval workflows for sensitive actions

Under The Hood

Architecture — The SDK’s src/agents/ package centers on agent.py (the Agent definition) and run.py/run_internal/ (the execution loop that resolves tools, guardrails, and handoffs each turn), with dedicated subpackages for tracing/, memory/ (session backends), mcp/ (Model Context Protocol client), voice/ and realtime/ (streaming audio agents), and sandbox/ (container-backed sandbox agents for platforms like E2B, Modal, Daytona, and Cloudflare). This separation lets the core agent loop stay provider-agnostic while extensions (extensions/) plug in specific model backends or tool ecosystems. Tech Stack — Pure Python 3.10+, built on openai (API client), pydantic v2 (schema/validation for structured outputs and function-call schemas), mcp for Model Context Protocol support, and websockets for realtime/voice streaming; a large matrix of optional extras (sqlalchemy, redis, dapr, temporal, mongodb, cloud sandbox providers) lets teams pull in only the session/sandbox backend they need. Code Quality — The repo carries 126+ files under tests/ covering agents, handoffs, guardrails, tracing, and sandbox providers, is fully typed (py.typed marker, pyright config), and is maintained with a very active commit cadence (health-score development-activity and maintenance-consistency both scored at the maximum). API Design — The primitives (Agent, Runner, handoff, function_tool) are deliberately small and composable, so a minimal single-agent script and a complex multi-agent handoff graph share the same mental model; instructions, tools, and guardrails are all passed as plain constructor arguments rather than requiring subclassing, keeping the barrier to a first working agent low.