Agents-OSS
EN
Submit

LangGraph vs OpenAI Agents SDK

Both are catalogued under Agent Frameworks. The figures come from the GitHub API; the assessments are ours.

LangGraphGraph-based runtime for stateful, controllable agentsOfficialMITOpenAI Agents SDKA deliberately small agent framework from OpenAIOfficialMIT

At a glance

At a glanceLangGraphOpenAI Agents SDK
LicenseMITMIT
LanguagesPython, TypeScriptPython
DeploymentSelf-hosted / Managed cloudSelf-hosted / Runs locally
MaturityEstablishedGrowing
Stars39.8k28.7k
Star growth over the last 7 days+14 ★+8 ★
Forks6.7k4.5k
Open issues69516
Last commit16 Aug 202616 Aug 2026
ActivityActiveActive

What each one does

LangGraph

Models an agent as an explicit graph of nodes and edges rather than a loop you cannot inspect. Durable checkpoints let a run pause for human approval and resume later, which is what makes it viable for workflows that touch production systems. The trade-off is verbosity: simple tool-calling agents need noticeably more setup here than elsewhere.

Full entry →

OpenAI Agents SDK

Few primitives — agents, handoffs, guardrails, tracing — and little else. The small surface area is the appeal: there is not much to learn and not much to fight. It is built around OpenAI's own models first, so weigh that if multi-provider portability matters to you.

Full entry →

What you can do

LangGraph

  • Resume a run after failure — Durable execution persists state as the graph advances, so an agent that dies mid-run restarts from exactly the step it left off rather than the beginning.
  • Pause for human approval — Interrupts halt execution at a chosen node so a person can inspect and modify the agent state before letting it continue.
  • Trace state transitions in LangSmith — Debugging is handed to LangSmith, a separate LangChain product paired with the graph: it visualises execution paths, captures state transitions and reports runtime metrics, so a wrong result is traced back to the node that produced it.
  • Start higher up with Deep Agents — Deep Agents is a package built on top of LangGraph for agents that plan, delegate to subagents and use a file system, saving you the graph wiring for those patterns.
  • Build the same graph in TypeScript — LangGraph.js is the equivalent library for JS/TS runtimes, so a Node service is not forced onto a Python sidecar.

OpenAI Agents SDK

  • Delegate between specialist agents — Handoffs and agents-as-tools let one Agent pass work to another, and Runner.run_sync executes the whole chain and returns final_output.
  • Give an agent a workspace — SandboxAgent takes a Manifest of entries such as GitRepo and runs commands, inspects files and applies patches through UnixLocalSandboxClient, or DockerSandboxClient on Windows.
  • Build voice agents two ways — RealtimeAgent holds a WebSocket session with gpt-realtime-2.1, while VoicePipeline chains speech-to-text, an agent workflow and text-to-speech from the optional voice extra.
  • Point it at models other than OpenAI's — The README calls the SDK provider-agnostic across the OpenAI Responses and Chat Completions APIs plus 100+ other LLMs, though every quickstart example still expects OPENAI_API_KEY in the environment.
  • Trace and resume conversations — Tracing is built in for viewing and debugging runs, and Sessions carries conversation history across separate Runner invocations without manual bookkeeping.

Other comparisons