Three primitives and a loop that runs until it's done
The official docs name three primitives: agents (an LLM with instructions and tools), agents as tools or handoffs, and guardrails. OpenAI's own pitch is a lightweight package with very few abstractions, a production-ready upgrade of its earlier Swarm experiment. Install is pip install openai-agents. Hello world is an Agent and Runner.run_sync. Nine lines, which is how most agent frameworks earn their first star. We haven't run it in production, so this is a teardown from the docs and comparison write-ups, not a war story.
The interesting part is the runner. The running agents page spells out the loop. Call the LLM. If the output is final, return it. If the model requests a handoff, swap the current agent and go again. If it produces tool calls, run them, append the results and go again. Exceed max_turns and you get a MaxTurnsExceeded exception.
That's all an agent is, and it's the part you'd otherwise write and debug yourself. Around it the SDK adds function tools generated from Python functions with Pydantic validation, MCP tool calling, sessions for memory, tracing, sandbox agents for isolated workspaces and realtime voice agents built on gpt-realtime-2.1.
Handoffs or agents as tools: pick one per boundary
The orchestration guide draws a clean line. With agents as tools, a manager agent keeps control and calls specialists through Agent.as_tool(). With handoffs, a triage agent routes the conversation and the specialist becomes the active agent for the rest of the turn.
Our opinion: start with agents as tools. One agent owns the final answer, and the docs say shared guardrails can be enforced in one place that way. Reach for handoffs when routing is itself the product, like a support desk where billing and technical questions need different instructions and different tone.
Guardrails are worth a mention because they're not a bolt-on. Per the docs they run in parallel with agent execution and fail fast when a check doesn't pass. That's a sensible default for input validation, though we'd still test what happens when a check trips halfway through a long tool call.
Durability is something you bring yourself
Here the SDK differs from the graph-first frameworks. AgentStack's comparison notes that LangGraph and Microsoft Agent Framework checkpoint natively, while the OpenAI Agents SDK leans on Temporal, DBOS, Restate or Dapr. The SDK docs confirm the integrations, and they come with different footprints: Restate needs its single-binary runtime, while DBOS needs only SQLite or Postgres.
On the human side, Developers Digest reports a serializable RunState at interruptions, with needs_approval and RunState.approve() for approval flows. The docs add a detail that matters: by default, tool input guardrails run only after approval, just before execution. You can set pre_approval_tool_input_guardrails=True to run them before the pause too.
If we were shipping this and already ran Postgres, we'd try DBOS first, because it adds the fewest moving parts. We can't tell you how any of these integrations behave under a real mid-run crash. None of our sources benchmarked them, and The Frontier Wire says it didn't either.
Where the bill and the lock-in hide
The SDK uses the Responses API by default for OpenAI models. Other providers go through ModelProvider, and Developers Digest lists the Any-LLM and LiteLLM adapters as beta. If one of those puts LiteLLM in your dependency tree, read our note on the LiteLLM supply-chain compromise first and pin versions.
None of our sources states a price for the SDK itself, and AgentStack lists no managed hosting for it: you host it. So the cost is tokens and your own infrastructure. A loop that continues until the task is complete means turns multiply, which is our reasoning, not a measured figure. Set max_turns deliberately instead of passing None.
Tracing is built in and defaults to OpenAI's own tooling, with 30+ external processors according to AgentStack. That's convenient. It also means you should decide early where traces, which contain prompts and tool outputs, are allowed to go. And with versions below 1.0 arriving every couple of weeks, budget time for upgrades.
Against LangGraph, Microsoft Agent Framework and OpenAI's own Agents API
The Frontier Wire ranks LangGraph and Microsoft Agent Framework first, then the OpenAI Agents SDK, Google ADK and Pydantic AI. That's one publication's ranking against its own criteria, with no benchmarks behind it. The reasoning is explicit control flow and native checkpoints. For the Microsoft side, see our teardown of Microsoft Agent Framework.
OpenAI also sells a managed option. Its agents guide puts the Agents API (a managed Codex runtime) at low integration effort, the Agents SDK at medium and the raw Responses API at high. The SDK is the middle path: it runs inside your application, and you own storage, approvals and deployment.
Developers Digest's advice matches ours: pick the vendor's own SDK when you're committed to that vendor. If you're multi-model, or you need a long-running graph with native checkpoints, look at LangGraph first.
Strengths, weaknesses, where it fits
Strengths
- Very few primitives, so the learning curve is short
- Handoffs and agents as tools cover most multi-agent shapes
- Guardrails run in parallel and fail fast
- Built-in tracing and MCP tool calling
Weaknesses
- Durable execution relies on Temporal, DBOS, Restate or Dapr
- Defaults to OpenAI models and OpenAI tracing
- Pre-1.0 versions arriving every few weeks
- No managed hosting for the SDK itself
Where it fits
- Lean multi-agent support or triage apps on OpenAI models
- Voice agents built on the realtime API
- Approval-gated tool use with human review
- Coding or document agents in isolated sandbox workspaces
Verdict: We'd put it in production for OpenAI-committed teams building lean agent apps, with a durable engine chosen and tested up front. We'd skip it for multi-model estates or long-running graph workflows that need native checkpoints.
Sources
- OpenAI Agents SDK · OpenAI
- Running agents - OpenAI Agents SDK · OpenAI
- Agent orchestration - OpenAI Agents SDK · OpenAI
- Agents | OpenAI API · OpenAI
- Best AI agent frameworks 2026: 10 open-source compared · The Frontier Wire
- Choosing an Agent Framework in Late 2026: 10 Frameworks Compared by Use Case · Developers Digest
- 12 Best AI Agent Frameworks in 2026, Compared for Production · AgentStack
Want this working in your stack, not just in a demo?
We are two senior engineers who build ai workflow automation with a fixed price per written scope.