← Back to research
·8 min read·opensource

OpenAI Agents SDK

The OpenAI Agents SDK is OpenAI's official agent framework — a lightweight, MIT-licensed successor to the Swarm experiment with 27,000+ GitHub stars, built on three primitives (Agents, Handoffs, Guardrails) plus built-in tracing, shipping weekly but still v0.x fifteen months after launch.

https://openai.github.io/openai-agents-python/ ↗
agent-frameworksopen-sourceopenaipythontypescriptmcp

Key takeaways

  • OpenAI's official agent framework — 27,000+ GitHub stars on the Python SDK plus 3,100+ on the JS/TS port, MIT-licensed, shipping roughly weekly releases (v0.17.5 on June 11, 2026) yet still pre-1.0 fifteen months after its March 2025 launch
  • Deliberately minimal: three core primitives (Agents, Handoffs, Guardrails) plus Sessions and a built-in agent loop, positioned as the production-ready upgrade of the Swarm experiment
  • Multi-model on paper — a LiteLLM extension covers 100+ providers and any ChatCompletions-compatible endpoint works — but the defaults (GPT models, Responses API, tracing to the OpenAI dashboard) all pull toward the OpenAI platform

FAQ

What is the OpenAI Agents SDK?

The OpenAI Agents SDK is OpenAI's official open-source framework for building multi-agent workflows, built around lightweight primitives — Agents, Handoffs, Guardrails, and Sessions — with built-in tracing.

How much does the OpenAI Agents SDK cost?

The framework is free and MIT-licensed. You pay for model API usage; the default tracing dashboard requires an OpenAI platform account.

Does the OpenAI Agents SDK work with non-OpenAI models?

Yes, with caveats. An official LiteLLM extension supports 100+ providers, and any ChatCompletions-compatible endpoint can be plugged in, but the Responses API, hosted tools, and default tracing are OpenAI-first, and some providers lack structured-output support.

Is the OpenAI Agents SDK the same as Swarm?

No. OpenAI describes the Agents SDK as the production-ready upgrade of Swarm, its earlier educational experiment in agent orchestration.

How is the OpenAI Agents SDK different from LangChain?

LangChain is a broad, provider-neutral ecosystem with 1000+ integrations and a commercial observability platform; the Agents SDK is a deliberately minimal, OpenAI-first framework with tracing built into the OpenAI dashboard.

Executive Summary

The OpenAI Agents SDK is OpenAI's official framework for building multi-agent workflows — a deliberately lightweight library built on three primitives (Agents, Handoffs, Guardrails) plus persistent Sessions, positioned as "a production-ready upgrade" of the Swarm experiment.[1] The Python repository, created in March 2025, stands at 27,000+ GitHub stars and 4,100+ forks as of June 2026, with an official JS/TS port at 3,100+ stars.[2][3]

The framework's design philosophy is "enough features to be worth using, but few enough primitives to make it quick to learn" — a direct contrast to heavier ecosystems.[1] It ships roughly weekly (v0.17.5 released June 11, 2026) but remains version 0.x fifteen months after launch, and its center of gravity is unmistakably the OpenAI platform: GPT models by default, the Responses API recommended, and tracing flowing to the OpenAI dashboard unless redirected.[2][4][5]

AttributeValue
CreatorOpenAI
ReleasedMarch 2025[2]
GitHub Stars27,000+ Python, 3,100+ JS/TS (as of June 2026)[2][3]
LicenseMIT[6]
Latest Versionv0.17.5 (June 11, 2026)[2]

Product Overview

The Agents SDK provides a built-in agent loop — call the LLM, run tools, loop until done — so developers compose agents from plain Python (or TypeScript) functions rather than learning a proprietary abstraction layer.[1]

Key Capabilities

CapabilityDescription
AgentsLLMs configured with instructions and tools
HandoffsAgents delegate work to other specialized agents
GuardrailsValidate agent inputs and outputs, fail fast on violations
SessionsPersistent conversation context across agent runs
Function toolsAny Python function becomes a tool with automatic schema generation
TracingBuilt-in workflow visualization, debugging, and monitoring
MCP integrationConnect MCP servers as tool sources
Voice/realtime agentsVoice agents with gpt-realtime-2 support[1]

Product Surfaces / Editions

SurfaceDescriptionAvailability
openai-agents (Python)Core framework, pip install openai-agentsGA since March 2025, v0.x[2]
openai-agents (JS/TS)Official port, includes voice agentsGA, v0.x[3]
Traces dashboardHosted trace viewer at platform.openai.com/tracesIncluded with OpenAI account[5]

Technical Architecture

The SDK is Python-first, leaning on native language features (decorators, type hints) for tool definition and orchestration rather than DSLs or config files.[1]

pip install openai-agents

Two model interfaces ship in the box: OpenAIResponsesModel (recommended, built on the Responses API) and OpenAIChatCompletionsModel for Chat Completions-style endpoints. The default model is gpt-5.4-mini as of June 2026, overridable per agent, per run, or via environment variable.[4] Non-OpenAI providers connect three ways: a global OpenAI-compatible client, a run-level ModelProvider, or the official LiteLLM extension (openai-agents[litellm]), which routes to 100+ providers via litellm/ prefixed model names.[7]

Key Technical Details

AspectDetail
DeploymentLibrary — runs anywhere Python/Node runs; no hosted runtime required
Model(s)OpenAI by default; 100+ providers via LiteLLM extension or any ChatCompletions-compatible endpoint[4][7]
ObservabilityTracing to OpenAI dashboard by default; 25+ external processors (Logfire, AgentOps, Braintrust, Langfuse, LangSmith, MLflow, Datadog)[5]
Open SourceYes (MIT License)[6]

Strengths

  • Minimal surface area — Three core primitives plus Sessions; the agent loop, schema generation, and handoffs work out of the box with very little ceremony[1]
  • First-party backing — Built and maintained by OpenAI, with roughly weekly releases (v0.17.5 shipped June 11, 2026)[2]
  • Tracing included free — End-to-end traces of LLM calls, tool runs, handoffs, and guardrails land in the OpenAI dashboard with zero setup[5]
  • Pluggable observability — 25+ documented external trace processors mean you are not locked into the OpenAI dashboard[5]
  • Both languages — Official Python and JS/TS SDKs, with voice/realtime agent support[3]
  • Community traction — 27,000+ stars and 4,100+ forks in fifteen months[2]

Cautions

  • OpenAI-first gravity — Multi-model support is real but second-class: the recommended Responses API is unsupported by many providers, hosted tools (file search, web search) are OpenAI-only, and the LiteLLM integration is labeled beta[4]
  • Tracing phones home by default — Traces upload to OpenAI's backend unless explicitly disabled or redirected; organizations with Zero Data Retention policies cannot use default tracing at all[5]
  • Still v0.x — Fifteen months after launch the SDK has not committed to a stable 1.0 API, and weekly releases mean ongoing churn[2]
  • Structured-output gaps off-platform — Some non-OpenAI providers lack JSON schema support, producing occasionally malformed outputs[4]
  • No commercial support tier — There is no enterprise edition, SLA, or paid support path; you get GitHub issues (137 open as of June 2026)[2]
  • Thin orchestration layer — No built-in durable execution, visual builder, or deployment platform; production infrastructure is your problem

Pricing & Licensing

TierPriceIncludes
Open SourceFreeFull Python and JS/TS SDKs (MIT License), built-in tracing

Licensing model: Open source (MIT); no commercial tier exists.[6]

Hidden costs: Model API usage (the SDK's defaults steer toward OpenAI models); the default traces dashboard requires an OpenAI platform account; production observability beyond the dashboard means wiring up a third-party processor with its own pricing.

Competitive Positioning

Direct Competitors

CompetitorDifferentiation
LangChain/LangGraphLangChain is provider-neutral with 1000+ integrations and a commercial platform; Agents SDK is minimal and OpenAI-first
CrewAICrewAI specializes in role-based multi-agent teams with an enterprise platform; Agents SDK is a lower-level primitive library
Pydantic AIPydantic AI is type-safety-first and provider-neutral; Agents SDK trades neutrality for first-party OpenAI integration
AutoGenAutoGen is Microsoft's research-oriented multi-agent framework; Agents SDK is production-minded and simpler

When to Choose OpenAI Agents SDK Over Alternatives

  • Choose OpenAI Agents SDK when: You are committed to OpenAI models and want the lightest official path to multi-agent workflows with free built-in tracing
  • Choose LangChain when: You need maximum provider neutrality, integration breadth, and a commercial observability platform
  • Choose CrewAI when: You want higher-level multi-agent team abstractions and enterprise support
  • Choose Pydantic AI when: Type safety and provider neutrality matter more than first-party OpenAI alignment

Ideal Customer Profile

Best fit:

  • Teams already standardized on OpenAI models and the OpenAI platform
  • Developers who want minimal abstractions — plain functions, decorators, a built-in loop
  • Projects that benefit from free, zero-setup tracing
  • Python or TypeScript teams building handoff-style multi-agent workflows
  • Builders of voice/realtime agents on OpenAI's stack

Poor fit:

  • Organizations requiring genuine multi-provider portability as a first-class concern
  • Teams with Zero Data Retention or strict data-residency requirements around default tracing
  • Enterprises that need commercial support, SLAs, or a managed runtime
  • Teams wanting high-level abstractions (crews, visual builders, deployment platforms)

Viability Assessment

FactorAssessment
Financial HealthBacked by OpenAI — no standalone business model, lives or dies by OpenAI's platform strategy
Market PositionStrong — 27,000+ stars in fifteen months, default choice for OpenAI-centric teams[2]
Innovation PaceRapid — roughly weekly releases, voice/realtime and MCP support added since launch[2]
Community/EcosystemHealthy — 4,100+ forks, 25+ third-party tracing integrations[2][5]
Long-term OutlookTied to OpenAI — abandonment risk is low, but roadmap serves OpenAI's platform interests first

The SDK's viability question is unusual: OpenAI is not going away, but the framework exists to make OpenAI's platform stickier, not to be a neutral standard. The persistent v0.x versioning suggests OpenAI prioritizes shipping over API stability guarantees — fine for fast-moving teams, a real cost for enterprises with long upgrade cycles.

Bottom Line

The OpenAI Agents SDK is the best-executed version of "less framework": three primitives, a built-in loop, free tracing, and weekly first-party releases. For teams already on OpenAI, it is the shortest path from idea to working multi-agent system, with less abstraction tax than any major alternative.

Recommended for: OpenAI-committed teams in Python or TypeScript who want lightweight multi-agent orchestration with built-in tracing and no platform fees.

Not recommended for: Organizations needing provider neutrality, commercial support, data-residency control over tracing, or a stable 1.0 API contract.

Outlook: The framework has earned real adoption — 27,000+ stars and a thriving processor ecosystem — but its gravity is the question.[2] Every default (models, Responses API, traces dashboard) points at the OpenAI platform, and the LiteLLM escape hatch is officially beta.[4][7] Watch for a 1.0 release as the signal that OpenAI considers this load-bearing infrastructure rather than a fast-moving companion library.

Research by Ry Walker Research • methodology

Sources

Share on XShare on LinkedIn