Three layers, one runtime

Ground school taught you the runtime. This course teaches the thing that sits on it.

The sentence to memorise

LangGraph is the graph runtime. LangChain's create_agent is a minimal agent harness on top of it. Deep Agents is a more opinionated harness on top of create_agent.

Three layers, strictly stacked. Not three competing products, not three ways to do the same thing. Each one is literally built out of the one below.

That has a consequence people get wrong constantly: the higher layers have no powers the lower ones lack. Deep Agents cannot do anything create_agent cannot be made to do, because Deep Agents is create_agent with a particular set of middleware already bolted on. The difference is assembly, not capability. Hold onto that, it comes back in lesson 3 with teeth.

What you already own

You know StateGraph, state schemas, reducers, nodes, edges and Command routing. Good, because create_agent returns a compiled StateGraph. Not a wrapper, not a lookalike. The literal thing you already know how to invoke, stream, checkpoint and rewind.

So every ground school reflex still fires. .invoke(), .stream(), a checkpointer, a thread_id, getState history. None of that changes. What changes is that you did not have to draw the graph.

What the exam actually looks like

Forty multiple choice questions, 120 minutes, pass at 28. Four domains at 25 percent each, which means Build is exactly 10 questions. Ten. That is the entire budget this course is fighting for.

It is semi open book: docs.langchain.com is consultable, general web search and AI are not. Think about what that does to question design. Nobody writes a 10-question section testing whether you can recall a parameter name that is one click away in a reference table.

So they test the things a docs page will not hand you in ten seconds: architecture, distinctions, execution order, what happens when you leave a default alone. Semantics over signatures. That is why this course spends a whole module on hook ordering and almost none on typing out function stubs.

There is a second consequence, and it is a study-time one. Anything genuinely lookup-able is worth less of your revision budget than it feels like it should be. Parameter spellings, optional field lists, the full catalogue of built-in middlewares: all one click away, all low yield. The high yield material is the stuff you would have to reason your way to under time pressure, which in this domain means ordering rules, defaults, and the boundaries between layers.

Every callout in this course is one of those. If you revise nothing else, revise the callouts.

How to read the rest of this course

Where something connects down into ground school, this course names the connection in one line and moves on. It will not re-teach reducers. If a lesson mentions a checkpointer and you feel a wobble, that wobble is a ground school problem, not a Build problem, and it has its own module over there.