There is a particular kind of confusion that the LangChain company has earned by naming three different products with the same four-letter prefix. People come to LangChain thinking it solves a problem, find it doesn’t quite, hear about LangGraph, assume it’s a different version of the same thing, try it, are still confused, and eventually discover that LangSmith is yet another product entirely. Once you understand that the three are not competing — that they sit at three different layers of the same stack — the confusion evaporates and the ecosystem starts to make sense.
This is the field guide that draws the line through all three. LangChain is the composition layer: a library for stitching together LLM calls, prompts, output parsers, retrievers, and tools into a single callable chain. LangGraph is the orchestration layer: when your chain stops being linear and starts being a stateful graph of nodes with branches and loops and human-in-the-loop pauses, LangGraph is what gives you the state machine. LangSmith is the observability layer: the trace, the eval harness, the dataset of failures you can replay, the dashboard that tells you which prompt change broke production.
Compose. Orchestrate. Observe. That is the whole picture, and once you have it, the rest of this guide is just filling in the details.
What it covers
Twelve sections, about twenty-two minutes if you read straight through.
§ 01 — The map in one breath. Compose → orchestrate → observe. The three-layer diagram that the rest of the guide is built on, with the question each layer answers.
§ 02 — LangChain — Build. The composition library. Prompts, models, output parsers, runnables, the LCEL pipe operator that turned LangChain v0.1 from an organizational mess into something coherent. What it’s good at (assembling LLM calls) and what it’s not (anything stateful).
§ 03 — LangGraph — Orchestrate. The graph runtime. Why you reach for LangGraph the moment your workflow has branches, cycles, or persistence requirements. State, checkpoints, the human-in-the-loop pause. Worked example of an agent loop expressed as a LangGraph.
§ 04 — LangSmith — Observe. The observability platform. Tracing, eval datasets, the human-feedback loop, the dashboard that turns “the model is misbehaving” into something diagnosable. The free tier, the pricing tier, where it stops being optional.
§ 05 — At a glance — which one, when. The decision table. A six-row matrix that maps the question you’re asking (“I need to add a step”, “my loops are getting hairy”, “I have no idea why this is failing in prod”) to the right tool.
§ 06 — The wider family. The other Lang products. LangServe — turn your chain into a FastAPI endpoint with one decorator. LangFlow — visual no-code builder on top of LangChain. LangFuse — open-source observability competitor (often paired with LangChain instead of LangSmith). The four-line summary of each.
§ 07 — Beyond the Lang stack. What the alternatives are. DSPy for declarative pipelines that auto-optimize. LlamaIndex for retrieval-heavy workloads. Pydantic AI for typed, validated tool-calling. Direct SDK for when the abstraction overhead isn’t earning its keep. When each one is the right reach.
§ 08 — When not to reach for them. The cases where these tools add friction without adding value. Simple one-shot prompts. Streaming-heavy UI. Workflows that already fit inside the model’s tool-calling. Where the abstraction tax is real and unhidden.
§ 09 — Talking to MCP. How LangChain and LangGraph interact with the Model Context Protocol. The integration story, where it’s tight, where it’s still rough. Cross-link to the API · CLI · MCP · ADK post for the broader picture.
§ 10 — All three in one workflow. A worked example. A document-Q&A system, built end to end: LangChain composes the retrieval-then-LLM pipeline, LangGraph adds a “did the answer cite a source? if not, try again” loop with a human-approval pause for sensitive topics, LangSmith captures every run and surfaces the prompt that produced the most hallucinations. The full code, narrated.
§ 11 — Glossary. Every term used in the guide, defined precisely. Chain. Runnable. Node. Edge. Checkpoint. Trace. Span. Dataset. Evaluator.
§ 12 — References & further reading. The relevant LangChain docs. The official LangGraph tutorials. The LangSmith pricing page. The handful of third-party write-ups that have actually changed my own practice.
Read it
The field guide lives at its own URL with a warm-paper editorial layout — teal-green for Build, terracotta for Orchestrate, gold for Observe (so the three tools color-code consistently throughout the document). Twelve sections, plain language, code examples where the abstractions earn them.
← Back to Autonomy