GILD verdict
Choose LangGraph when code owns the path and runs resume from saved state, and the Agents SDK for OpenAI-model loops with handoffs.
- The workflow
- a team choosing the runtime for agents that call tools, wait for people and must survive a restart
- The decisive evidence
- Reported3 of 5 named teams adopted one; What teams report below sets out each account.
- DocumentedThe Agents SDK keeps sessions but leaves durable execution to Temporal, Restate, DBOS or Dapr, and Temporal's integration has been generally available since Mar 23, 2026.523
- GILD analysisNo named team has published a production account of the Agents SDK on its own blog that GILD found, so how it holds up in production against LangGraph is unresolved. Self-hosted, GILD's calculation puts 60,000 requests a month at US$1,656 on either.
- The tradeoff
- LangGraph also lets several model providers share one runtime; the Agents SDK brings built-in tracing. LangGraph gives you checkpoints and interrupts in one library at the cost of more code and checkpoint housekeeping, and the Agents SDK needs less code but a second engine for durability and a pinned minor version, since its 0.x minors can break interfaces.1114202526
- What would change the call
- The call moves when a named team publishes a production Agents SDK account with its own measurements, or when the SDK reaches 1.0 with built-in checkpointing.
The accounts on record split the pair by who owns durability. GitLab, Wayfair and LinkedIn run production agents on LangGraph's checkpointed graphs, and builders who self-host it add their own worker fencing. The OpenAI Agents SDK keeps sessions but no checkpoints, so a crashed long run restarts unless you add Temporal or another engine, and a team tracing it to LangSmith saw duplicate traces. Aurelio Labs used both and moved to its own framework. Cost does not separate them when self-hosted, since model tokens dominate the bill.
What teams report
Failures and incidents 1
Karolis Mariunas's teamOn OpenAI Agents SDK
Tracing an Agents SDK app with handoffs to LangSmith produced duplicate traces with different ids, which broke the team's dashboard aggregations (GitHub issue, Mar 31, 2025).
“This makes our dashboards kind of useless, because we cannot trust the aggregations and metrics…”
Karolis Mariunas, Developer running the Agents SDK with LangSmith tracing, Karolis Mariunas's teamGitHub, langchain-ai/langsmith-sdk, , post
Left or turned down 1
Aurelio Labs
Briggs, who maintains Aurelio Labs' competing graphai framework, wrote on Hacker News that his team had used LangChain, LangGraph and OpenAI's Agents SDK extensively, then left all three for its own async framework, which he calls a much better and cleaner experience (Oct 8, 2025).
“…s agents SDK pretty extensively. Now we roll our own, generally a much better and cleaner experience.”
James Briggs (jamesbriggs), Maintainer of graphai, Aurelio Labs' competing agent framework, Aurelio LabsHacker News, , post
Adopted 3
GitLabOn LangGraph
GitLab researched agent frameworks and chose LangGraph, for durable execution and traceability, as the base of Duo Agent Platform (Sep 22, 2026).
“We researched available AI frameworks and selected LangGraph, an agent runtime and low-level orchestration framework from LangChain, as the foundation for GitLab Duo Agent Platform.”
Mikolaj Wawrzyniak, Engineer, GitLab Duo Agent Platform, GitLabGitLab blog, , write-up WayfairOn LangGraph
Wayfair runs its Wilma ticket flagger on an event-driven LangGraph pipeline, at 95% precision and 97% recall against human targets of 80% and 90% (Jan 15, 2026).
“The STP Intervention Flagging system runs on an event-driven automation pipeline using LangGraph.”
Farshid Bahrami, Co-author, Wayfair Tech Blog, WayfairWayfair Tech Blog, , write-up TemporalOn OpenAI Agents SDK
Temporal built an integration that adds durable execution to Agents SDK agents through its Python SDK, generally available from Mar 23, 2026, which fills the checkpointing the SDK leaves out.
“Update: As of March 23rd, 2026, the integration between the OpenAI Agents SDK and Temporal's Python SDK is now Generally Available.”
Temporal, Durable execution vendor and integration partner, TemporalTemporal blog, , write-up
Choosing between LangGraph and OpenAI Agents SDK
Choose LangGraph if
- Your code should decide the order of steps, with branches and retries you can read, as GitLab's Duo Agent Platform flows and Wayfair's ticket flagger run.
- Runs must survive restarts and resume from the last step without adding a separate workflow engine; independent comparisons name this as the gap on the Agents SDK side (Aug 9, 2026).
- You use several model providers, or want a stable 1.x line with breaking changes only in major releases.
Choose OpenAI Agents SDK if
- Your agents run mostly on OpenAI models and you want handoffs, guardrails, and sessions with little code; an independent comparison calls it a close second when you only use OpenAI models (Oct 2, 2026).
- You want tracing in the OpenAI dashboard without setting up another platform.
- You already run Temporal, Restate, DBOS, or Dapr and want durability from that engine; Temporal's integration has been generally available since Mar 23, 2026.
Side by side
| Dimension | LangGraph | OpenAI Agents SDK |
|---|---|---|
| What teams reportAurelio Labs used both extensively before moving to its own framework (Oct 8, 2025). | GitLab chose LangGraph as the base of Duo Agent Platform (Sep 22, 2026), Wayfair's LangGraph ticket flagger reached 95% precision and 97% recall (Jan 15, 2026), and LinkedIn adapted LangGraph to its messaging and memory infrastructure (Sep 10, 2025).123 | Temporal shipped a durable-execution integration for the Agents SDK, generally available Mar 23, 2026. A developer tracing an Agents SDK app to LangSmith got duplicate traces that made the team's dashboards untrustworthy (Mar 31, 2025).56 |
| Design | A runtime: an explicit graph, or the Functional API, where your code defines each transition | A framework: agents with instructions and tools, handoffs or agents as tools for delegation, and guardrails, in a loop the model drives |
| License and versioningLatest releases on each project's GitHub repository. | MIT, free; Python (langgraph 1.2.18, Oct 26, 2026) and JavaScript (@langchain/langgraph 1.5.3, Oct 26, 2026). 1.x is an LTS line: breaking changes only in a major release, and 1.0 maintained for at least a year after 2.0 | MIT, free; Python (openai-agents 0.24.0, Oct 15, 2026) and TypeScript (@openai/agents 0.21.0, Oct 20, 2026); 0.Y.Z versioning in which a minor release may break public interfaces |
| Models | Any provider, through LangChain chat models or the provider's own SDK inside a node | Responses API by default for OpenAI models; other providers through Chat Completions-compatible endpoints or the LiteLLM and Any-LLM adapters |
| State and durability | Checkpointers save state after each step to memory, SQLite, Postgres, or MongoDB, with exit, async, and sync durability modes; threads, a cross-thread store, replay, and fork | Sessions keep conversation history in SQLite, Redis, SQLAlchemy, MongoDB, Dapr, encrypted, or OpenAI-hosted stores; durable execution through Temporal, Restate, DBOS, or Dapr integrations |
| Human review | interrupt() pauses any node with state persisted; Command(resume=...) continues later, in another process if needed | Tools that need approval pause the run; RunState serializes it so it can resume after a decision |
| Tracing | Off unless LANGSMITH_TRACING is set; Studio for local runs; stream modes for your own tooling | On by default, to OpenAI's traces dashboard; off with OPENAI_AGENTS_DISABLE_TRACING; unavailable to organizations under Zero Data Retention |
| HostingList prices in US dollars. | Your own services, or LangSmith Deployment: Plus at US$39 per seat per month with 1 free Small Serverless deployment, then compute metered in LSUs at US$1 each; Enterprise adds hybrid and self-hosted | A library you host yourself; costs are the model and tool usage it calls |
| 60,000 requests a month, 10 peopleGILD's calculation from published rates, Oct 28, 2026. Inputs: 60,000 requests a month, each with 10,000 prompt tokens (2,000 uncached, 1,000 cache writes, 7,000 cache reads) and 2,000 output tokens on Claude Sonnet 5.5 at US$2 and US$10 per million; one DigitalOcean droplet at US$24; 10 people, one LangSmith seat. Operations time is excluded. | US$1,656 a month self-hosted: US$1,632 in model tokens and US$24 for one 2 vCPU, 4 GiB server. On LangSmith Plus with one Small Dedicated deployment always on, US$2,055.12: one US$39 seat and US$384.12 in compute at US$1 per LSU.7819 | US$1,656 a month self-hosted, the same tokens and server; the SDK adds no fee and has no managed hosting to price.7827 |
| GitHub starsCounted Oct 12, 2026. Stars count attention and say nothing about production use. | 43,144 (Python) and 3,350 (JavaScript) | 29,988 (Python) and 3,910 (JavaScript) |
What practitioners report, side by side
Each product's figures come from its own practitioners, for the group you pick.
Overall signal
All practitioners rate LangGraph +25, 42 points above OpenAI Agents SDK at −17. Most of the gap is in complaints, with 21 of OpenAI Agents SDK's 50 sources negative against 8 of LangGraph's 62.
The score counts the authors from all practitioners that take a side, one vote each with their latest view, and is how far positive votes lead negative ones, per 100 votes, so LangGraph's +25 rests on votes split 23 positive, 30 mixed and 8 negative and OpenAI Agents SDK's −17 on 11, 16 and 19.
A lead of a fifth of the votes reads positive and a lead of half strong positive, with the same lines below zero. The dots show each source, positive, mixed, negative. Build one pilot agent to measure the glue code and debugging time the framework adds.
Footprint on the AI Application Platform stack
They differ on three of 9 layers. Coverage is each product's placement on this stack, as its product page shows it. Core is a job the product owns; Supports and Touches are jobs it helps with. Coverage says nothing about how well a product does them.
| Layer | LangGraph | OpenAI Agents SDK |
|---|---|---|
| Operate | ||
| Observability | Touches | Supports |
| Cloud deployment | Supports | Touches |
| Guardrails | Touches | Supports |
The 6 layers where they match
| Layer | LangGraph | OpenAI Agents SDK |
|---|---|---|
| Model | ||
| Model APIs | Touches | Touches |
| Model gateway | Touches | Touches |
| Build | ||
| Orchestration | Core | Core |
| Tool use | Supports | Supports |
| Data | ||
| Retrieval | Touches | Touches |
| Evaluate | ||
| Evaluation | Touches | Touches |
What practitioners praise
Among all practitioners, LangGraph alone has four praise themes that reach 5 sources from 3 authors. Explicit Control Flow and Checkpointing & Durable State lead, at 22 of 25 and 10 of 18. A Small Set of Primitives and Agents Built and Shipped reach 5 sources from 3 authors only for OpenAI Agents SDK, at 11 of 11 and 9 of 9. LangGraph needs a pilot on the same agent task before you choose on either.
| Praised for | LangGraph | OpenAI Agents SDK |
|---|---|---|
| Explicit Control Flow | 22 of 25 | Not a theme in its sources |
| Checkpointing & Durable State | 10 of 18 | Not a theme in its sources |
| Production Readiness | 9 of 15 | Not a theme in its sources |
| Human-in-the-Loop & Time Travel | 4 of 5 | Not a theme in its sources |
| A Small Set of Primitives | Not a theme in its sources | 11 of 11 |
| Agents Built and Shipped | Not a theme in its sources | 9 of 9 |
Each count is the sources that praise a theme, or raise it as a concern, out of those that discuss it. A row pairs the two products when both carry a theme under the same or an equivalent name.
A count ending in "so far" has under 5 sources from 3 authors among all practitioners. "Not a theme in its sources" means GILD found no theme of that name in that product's sources.
What practitioners raise as concerns
Among all practitioners, OpenAI Agents SDK alone has five concern themes that reach 5 sources from 3 authors. Runtime Defects in Production and Latency & Run Overhead lead, at 14 of 14 and 7 of 7. LangGraph alone has three concern themes that reach 5 sources from 3 authors. Complexity & Overhead and Breaking Changes & Docs lead, at 10 of 13 and 9 of 10.
| Raised as a concern | LangGraph | OpenAI Agents SDK |
|---|---|---|
| Complexity & Overhead | 10 of 13 | Not a theme in its sources |
| Breaking Changes & Docs | 9 of 10 | Not a theme in its sources |
| Typing & Language Support | 4 of 5 | Not a theme in its sources |
| Runtime Defects in Production | Not a theme in its sources | 14 of 14 |
| Latency & Run Overhead | Not a theme in its sources | 7 of 7 |
| OpenAI-First Design & Adapter Gaps | Not a theme in its sources | 7 of 8 |
| Hosting & Storage Left to the Team | Not a theme in its sources | 5 of 5 |
| Tracing Defaults & Export Gaps | Not a theme in its sources | 4 of 5 |
27 sources
- GitLab blog: How GitLab reduced code-per-agentic-flow ratio by 45% (Sep 22, 2026)
- Farshid Bahrami and colleagues (Wayfair Tech Blog): Flag Early, Fix Faster: Real-Time Intervention for Supplier Transfers with Wilma (Jan 15, 2026)
- Karthik Ramgopal and colleagues (LinkedIn Engineering Blog): The LinkedIn Generative AI Application Tech Stack, extending to build AI agents (Sep 10, 2025)
- Hacker News comment on moving from LangGraph and the Agents SDK to an in-house framework (Oct 8, 2025)
- Temporal blog: Announcing the OpenAI Agents SDK integration, generally available Mar 23, 2026 (Jul 30, 2025)
- GitHub langchain-ai/langsmith-sdk#1619: duplicate traces through the OpenAI Agents SDK (Mar 31, 2025)
- Claude API pricing: Claude Sonnet 5.5 token rates (checked Oct 28, 2026)
- DigitalOcean Droplet pricing: 2 vCPU, 4 GiB at US$24 a month (checked Oct 28, 2026)
- LangGraph docs: Overview
- LangChain docs: Runtimes, frameworks, and harnesses
- LangGraph docs: Checkpointers and durability modes
- LangGraph docs: Interrupts
- LangGraph docs: Studio
- LangChain docs: Release policy (LangGraph LTS)
- LangGraph on GitHub (MIT license, stars, issues)
- LangGraph release 1.2.17 on GitHub (Oct 24, 2026)
- LangGraph release 1.2.18 on GitHub (Oct 26, 2026)
- @langchain/langgraph on npm (release history)
- LangSmith plans and pricing (LangSmith Deployment rates)
- OpenAI Agents SDK docs: Overview
- OpenAI Agents SDK docs: Models (non-OpenAI providers)
- OpenAI Agents SDK docs: Sessions
- OpenAI Agents SDK docs: Running agents (durable execution integrations)
- OpenAI Agents SDK docs: Human in the loop
- OpenAI Agents SDK docs: Tracing
- OpenAI Agents SDK docs: Release process (0.Y.Z versioning)
- OpenAI Agents SDK on GitHub (MIT license)