What it is
The OpenAI Agents SDK is OpenAI's open-source Python library for agents and multi-agent workflows. Its primitives are agents (a model with instructions, tools, guardrails and handoffs), handoffs and agents-as-tools for delegation, guardrails on input, output and tool calls, sessions for conversation history, human-in-the-loop approvals and built-in tracing.
Since an April 2026 update it also has sandbox agents, which work in a workspace with files, shell commands and patches over long tasks. That workspace can be the local machine, Docker, or a hosted provider: E2B, Modal, Daytona, Cloudflare, Vercel and Blaxel all have clients. Realtime voice agents and a speech pipeline round it out, and a TypeScript version lives in a separate repo.
It calls OpenAI's Responses or Chat Completions APIs by default and reaches other providers through LiteLLM or any-llm adapters. It needs Python 3.10+; version 0.24.0 came out on 15 October 2026 and the project is still pre-1.0.
Our review
The Agents SDK is the shortest route from an OpenAI API key to a working multi-agent system, and its defaults are the part to read before shipping. Tracing is on out of the box and exports every model call, tool call, handoff and guardrail to OpenAI's Traces dashboard. You turn it off with OPENAI_AGENTS_DISABLE_TRACING=1 or per run, and it is unavailable to organisations under Zero Data Retention: useful in development, a data-flow question in a security review.
The core is small. An Agent plus Runner.run_sync is a working program; handoffs and agents-as-tools cover most delegation patterns, and guardrails can block a request before the agent starts or check individual tool inputs. The repo has about 30,200 stars and shipped eight releases between 16 August and 15 October 2026, from 0.21.1 to 0.24.0, so expect churn while it stays below 1.0.
Sandbox agents move it towards a production harness, and the docs are frank about their edges. The default UnixLocalSandboxClient runs commands as ordinary host processes: on Linux it adds no OS-level confinement, on macOS sandbox-exec restricts files but not network, and every command inherits the full host environment, API keys included, unless you set inherit_host_environment=False. Untrusted input belongs in the Docker client, where network_mode="none" cuts networking, or with a hosted provider.
Provider neutrality is real but second-class. Adapters reach other models, yet hosted tools and the tracing dashboard are OpenAI services, and new primitives arrive tuned for OpenAI models such as gpt-realtime-2.1. Pydantic AI and LangGraph make fewer assumptions about the vendor, and LangGraph gives explicit graph control flow that this SDK leaves to handoffs and plain Python. Outside contributors cannot send pull requests at all; the repo accepts them only from collaborators.
Where it earns its place
A backend engineer at a fintech start-up already on the OpenAI API
Route support chats between a triage agent and billing and refunds specialists
Define three agents with handoffs and an input guardrail that rejects out-of-scope requests before the run starts; traces show each hand-off. The SDK is free under MIT, so the bill is tokens, and each hand-off adds model calls. Disable tracing or check the data agreement before customer data flows through it.
A platform team building an internal code-review bot
Let an agent clone a repo, run tests and propose a patch
A SandboxAgent with a GitRepo manifest entry and the DockerSandboxClient (openai-agents[docker]) keeps commands in a container; network_mode="none" isolates untrusted code. Moving to E2B, Modal or Daytona later changes the run config, not the agent definition.
A product manager at a 50-person SaaS company scoping a voice assistant
Prototype a spoken support agent without stitching speech services together
RealtimeAgent runs over WebSocket with gpt-realtime-2.1, or VoicePipeline chains speech-to-text, the agent and text-to-speech. The prototype reuses the same tools and guardrails as the text agents, but audio tokens become the main cost and the design is tied to OpenAI's realtime models.
Before you commit
- Setup
- `pip install openai-agents` (or `uv add openai-agents`) on Python 3.10+, set OPENAI_API_KEY, and the README's first agent runs in minutes. Extras: [voice], [redis] for sessions, [docker] for container sandboxes.
- Integrations
- OpenAI Responses and Chat Completions by default; other models through LiteLLM or any-llm; function tools, MCP servers and hosted tools; sandbox clients for E2B, Modal, Daytona, Cloudflare, Vercel and Blaxel; custom trace processors to send traces elsewhere.
- Data and privacy
- Tracing is on by default and exports run data to OpenAI; it is unavailable under Zero Data Retention and can be disabled globally or per run. The Unix-local sandbox passes the full host environment to commands by default and gives no confinement on Linux. MIT licence.
- Cost at scale
- The SDK is free; production cost is model tokens and hosted tool calls at OpenAI's API prices, plus a sandbox provider's compute if you use one. Multi-agent designs multiply calls per user request, so spend grows with the number of agents a request passes through.
- Limits
- Pre-1.0 with frequent releases; sandbox features launched Python-first; no graph-style control flow; contributions limited to collaborators.
What's good
- An Agent and a Runner make a working program
- Handoffs, guardrails and approvals built in
- Sandbox clients for Docker, E2B, Modal, Daytona and more
- Tracing on by default, with custom processors for other back ends
What's not
- Pre-1.0, with eight releases in two months
- Tracing sends run data to OpenAI unless disabled
- Local sandbox inherits host env and has no Linux confinement
- Pull requests only from collaborators
Best for
Developers building agents mainly on OpenAI models who want a minimal, well-documented SDK.
Alternatives
OpenAI Agents SDK: common questions
- Is OpenAI Agents SDK worth it?
- A small, readable agent SDK whose defaults quietly assume OpenAI: tracing ships your runs to its dashboard, and the local sandbox trusts your machine. We rate it 4.4 out of 5.
- How much does OpenAI Agents SDK cost?
- Free and open source under the MIT license. You pay for model usage.
- What are the downsides of OpenAI Agents SDK?
- Pre-1.0, with eight releases in two months; tracing sends run data to OpenAI unless disabled; local sandbox inherits host env and has no Linux confinement; pull requests only from collaborators.
- Who is OpenAI Agents SDK best for?
- Developers building agents mainly on OpenAI models who want a minimal, well-documented SDK.
- How much does OpenAI Agents SDK cost for a team or in production?
- The SDK is free; production cost is model tokens and hosted tool calls at OpenAI's API prices, plus a sandbox provider's compute if you use one. Multi-agent designs multiply calls per user request, so spend grows with the number of agents a request passes through.
- Is OpenAI Agents SDK safe for company data?
- Tracing is on by default and exports run data to OpenAI; it is unavailable under Zero Data Retention and can be disabled globally or per run. The Unix-local sandbox passes the full host environment to commands by default and gives no confinement on Linux. MIT licence.
- What are the best OpenAI Agents SDK alternatives?
- The closest alternatives we have reviewed are Pydantic AI, LangGraph, Google Agent Development Kit (ADK) and Microsoft Agent Framework.
Answers checked Oct 23, 2026.
Make OpenAI Agents SDK? Show this rating on your site
