LangChain, LangGraph, Deep Agents, LangSmith

The same job, four runtimes.

Take one job and run it as a chain, a LangChain agent, a Deep Agent, or a LangGraph. See where each one fits, and where it stops.

The stack

Deep Agents sits on LangChain. Both sit on LangGraph. LangSmith is how you watch a run.

Select a layer. The description stays until you pick another.

create_deep_agent adds a filesystem, skills, subagents, and planning on top of the agent loop.

The open-source stack

chain

A pipeline with a known order. langchain.com barely mentions this anymore, but it is still the right tool when you already know the steps.

loop

LangChain's create_agent: a model in a loop with tools. Built on LangGraph, without making you draw the graph.

harness

Deep Agents: the same loop, plus a filesystem, skills, subagents, and planning, already wired up.

graph

LangGraph is the runtime underneath: durable runs, streaming, human-in-the-loop. Learn LangGraph is the graph practice.

Where to start

Pick a job and a runtime, then step through what happens. After a run, the Smith panel shows a fake trace.

Five situations, each with a runtime that fits and one that does not.

Glossary

LangChain's site uses these as product names. Here they are just the machines.

Chain: A pipeline. You already know the steps and the order. langchain.com barely talks about this now; it is still the right tool for a one-pager.
LangChain: create_agent: a model in a loop with tools. The framework. You do not have to draw a graph to start.
Deep Agents: create_deep_agent: that same loop, plus a filesystem, skills, subagents, and planning, already attached.
LangGraph: The runtime underneath. Durable runs, streaming, human-in-the-loop. You draw the graph when the loop is not the right shape. Practice that in Learn LangGraph. Open Learn LangGraph
LangSmith: Traces, evals, and deploy. How you watch a run after it leaves your laptop.
Agent Server: LangSmith's host for an agent: endpoints, sandboxes, a place for a schedule.
Fleet: The no-code builder. For people who are not going to open a repo.
Engine: Groups similar production failures and can suggest a fix. You still have to accept it.
LLM Gateway: Sits in front of the models: spend caps, PII, routing.

FAQ

Do I need an API key?

No. The studio is simulated. Nothing calls a model.

Is this official LangChain?

No. A community resource from En Dash. The product names are real; the runs are not.

When is a chain enough?

When the steps are known and the work never comes back. The PDF one-pager is a chain. The rental that waits for a person is a graph.