Werkzeuge
·
5 MIN
LangChain, LangGraph, LangSmith, LangFlow: four names, two companies, one decision
Three of these four are one converging stack. LangFlow is a different company. Pick by the job, not the brand.
LOCATION
Worldwide
SERIES
AI Tooling & Dev Harnesses
AUTHOR
Sayan Sinha
PUBLISHED
The four names are not a suite you adopt together. Three of them are one company's converging stack, and the fourth belongs to a different company entirely. LangChain, LangGraph, and LangSmith are all built by LangChain Inc, and in 2025 they deliberately merged into a single agent platform. LangFlow is a separate visual builder, originally from DataStax and now part of IBM. The useful 2025 to 2026 story is consolidation on one side and a security wake-up on the other. The right move is to decide by the job in front of you, not by adopting four tools because their names rhyme.
01.
What actually shipped in 2025 and 2026
The headline is stability after years of churn. LangChain 1.0 and LangGraph 1.0 both reached stable release on 22 October 2025, with a commitment to no breaking changes until 2.0. The framework's new headline abstraction, create_agent, now runs on the LangGraph runtime underneath, so the agent loop is a LangGraph graph. That is the clearest signal of the merge: the framework and the orchestration runtime are no longer separate choices.
LangGraph 1.0 brought durable execution as a built-in: state is checkpointed automatically, so a long-running or human-in-the-loop workflow resumes after an interruption or a server restart. LangSmith, the observability and evaluation layer, added full OpenTelemetry support in March 2026, which means you can send traces to it from code that does not use LangChain at all. The commercial picture matched the technical one: LangChain Inc raised a 125 million dollar Series B at a 1.25 billion valuation in October 2025, and renamed LangGraph Platform to LangSmith Deployment, folding deployment, observability, and evaluation into one product.
02.
Four names, two companies
Sorting out who makes what is the single most useful thing before you choose. Most write-ups blur it; getting it right changes the decision.
Tool | Owner | What it is | When it earns its place |
|---|---|---|---|
LangChain | LangChain Inc | The framework and agent abstractions | You want broad model and tool integrations behind one interface |
LangGraph | LangChain Inc | Stateful orchestration runtime | You need durable, resumable, human-in-the-loop agents |
LangSmith | LangChain Inc | Observability, tracing, evaluation | You need to see and test what your agents actually do |
LangFlow | DataStax, now IBM | Visual low-code flow builder | Non-engineers need to prototype flows on a canvas |
LangFlow is the one that trips people up. It is not a LangChain Inc product. It is an open-source visual builder created by DataStax, whose parent was acquired by IBM in February 2025, and it now points toward IBM's watsonx. The name rhyme is a coincidence. If your requirement is a drag-and-drop canvas for people who do not write Python, LangFlow is the candidate; it is not a layer of the LangChain stack.
03.
The decision most teams actually face
Once the names are clear, the choice is usually a short decision tree, and you rarely need all four.
Need durable orchestration? LangGraph is the strongest piece, and the one even critics tend to respect. Resumable state, checkpointing, and human-in-the-loop are real engineering, not wrappers.
Need observability and evaluation? LangSmith is capable, but it couples you to the ecosystem. Langfuse is the open-source, self-hostable counter-pressure, and the most-cited escape hatch. Now that LangSmith speaks OpenTelemetry, you can trace from non-LangChain code, which softens the lock-in.
Building a straightforward agent? The provider SDKs (the OpenAI Agents SDK, the Claude Agent SDK) and Pydantic AI are credible, lighter alternatives. The dominant practitioner gripe is the layers of abstraction in the framework, and several teams report cutting code and maintenance by moving to a thinner stack.
Non-engineers prototyping? LangFlow, with the caveat that a canvas prototype still needs the same governance and evaluation before it ships.
This is the same lesson we drew when comparing how to wire AI into a stack through an API, an MCP server, or a CLI: the tools are layers, and you pick by the job, not the novelty. It also rhymes with what we found looking at no-code agent builders for the Mittelstand, where the canvas gets you a demo fast and stops exactly at governance, reliability, and cost.
04.
Stable did not mean hardened
One caveat matters for a business audience: 1.0 bought API stability, not a hardened security posture. A cluster of CVEs landed against LangChain and LangGraph in late 2025 and early 2026, led by a deserialization flaw that could leak API keys and environment secrets, rated CVSS 9.3 and nicknamed LangGrinch. Two more followed: a path-traversal issue in the prompt-loading API, and a SQL-injection in the LangGraph SQLite checkpointer.
The practical reading is not to avoid the framework. It is to treat it like any other dependency that touches your secrets and your data: pin versions, apply the patched releases, and include it in your threat model. A framework that orchestrates tool calls and holds credentials is part of your attack surface, and "stable" on the version number says little about that. For a regulated or on-prem deployment, that review belongs in the same conversation as data residency, not after go-live.
