31 Phases to Learn LangChain & LangGraph
If you've been following @prashant.code on Instagram, you've probably seen the reel that maps out the entire path to learning LangChain and LangGraph — from the absolute basics all the way to fully agentic, production-grade systems. This post is the full written version of that roadmap, with a short explanation of what's actually happening at every phase.
There isn't really a shortcut through this. Jumping straight to "agents" without understanding models, prompts, and retrieval first is exactly how people end up debugging systems they don't actually understand. Follow it in order.
Foundations
1LLM fundamentals- 2LangChain models
- 3Prompt templates
- 4Runnables / LCEL
- 5Structured output
Data & Retrieval
6Document loaders- 7Text splitting
- 8Retrievers
- 9Advanced RAG
Tools & Agents
10Tools- 11Tool calling
- 12Agents
LangGraph Core
13LangGraph State- 14Nodes
- 15Edges
- 16Conditional routing
- 17Loops
- 18Persistence
Runtime & Execution
19Memory- 20Human-in-the-loop
- 21Streaming
- 22Async / parallel execution
- 23Error handling
Advanced Systems
24Agentic RAG- 25Multi-agent systems
- 26LangSmith
Production & Safety
27Evaluation- 28Security / guardrails
- 29MCP
- 30Production architecture
Mastery
31Advanced agentic systems
What each phase actually means
The 31 phases aren't random — they fall into six themes, each one building on the last. Foundations teach you how a model thinks. Data & Retrieval teaches it your own knowledge. Tools & Agents give it hands. LangGraph Core gives those hands a nervous system. Runtime & Execution makes it reliable enough to run unattended. And Advanced Systems / Production & Safety is where a personal project turns into something you'd actually trust in front of real users.
Foundations
- 1
LLM fundamentals
What a large language model actually does under the hood: tokens, context windows, and why “predicting the next token” is the whole game.
- 2
LangChain models
How LangChain wraps different LLM providers (OpenAI, Anthropic, local models) behind one consistent interface, so you can swap models without rewriting your app.
- 3
Prompt templates
Turning prompts into reusable, parameterized templates instead of hardcoded strings scattered through your code.
- 4
Runnables / LCEL
LangChain's composition language for chaining steps together with `|`, so a chain reads like a pipeline instead of a pile of function calls.
- 5
Structured output
Forcing the model to return data in a shape your code can actually use (JSON, Pydantic models) instead of parsing free-form text.
Data & Retrieval
- 6
Document loaders
Pulling raw content in from PDFs, websites, Notion docs, and databases — the first step of getting your own data into an LLM pipeline.
- 7
Text splitting
Breaking large documents into chunks small enough to embed and retrieve, without cutting sentences in half and losing meaning.
- 8
Retrievers
The component that takes a query and returns the most relevant chunks from your data — the “search” half of RAG.
- 9
Advanced RAG
Techniques beyond naive retrieval: re-ranking, hybrid search, and query rewriting to actually get accurate answers instead of “close enough.”
Tools & Agents
- 10
Tools
Giving an LLM the ability to call real functions — search the web, query a database, hit an API — instead of just generating text.
- 11
Tool calling
The actual mechanism: how a model decides which tool to call, with what arguments, and how that call gets executed safely.
- 12
Agents
LLMs that don't just answer once, but reason, decide on an action, observe the result, and loop until the task is actually done.
LangGraph Core
- 13
LangGraph State
The shared object that flows through every step of a graph — effectively the “memory” of what's happened so far in a run.
- 14
Nodes
Individual units of work in a LangGraph graph: a node reads state, does something, and returns updated state.
- 15
Edges
The connections that decide which node runs next, turning a set of nodes into an actual flow instead of a disconnected pile of functions.
- 16
Conditional routing
Edges that branch based on state — if the model needs a tool, go here; if it's done, go there.
- 17
Loops
Letting a graph return to an earlier node, which is how agents “keep working” instead of running once and stopping.
- 18
Persistence
Saving graph state so a run can pause, resume, or recover from a crash without starting over from scratch.
Runtime & Execution
- 19
Memory
Giving an agent recall across turns or sessions, from simple chat history to long-term semantic memory.
- 20
Human-in-the-loop
Pausing a graph to ask a person for approval or input before continuing — critical for anything with real-world consequences.
- 21
Streaming
Sending tokens, state updates, or events back to the user as they happen, instead of making them wait for the whole run to finish.
- 22
Async / parallel execution
Running independent steps concurrently instead of one after another, so multi-step agents don't crawl.
- 23
Error handling
What happens when a tool fails, a model times out, or output doesn't parse — the difference between a demo and something production-ready.
Advanced Systems
- 24
Agentic RAG
Retrieval that isn't a single fixed step, but something the agent decides to do, redo, or skip based on what it's learned so far.
- 25
Multi-agent systems
Multiple specialized agents collaborating, or handing off to each other, instead of one agent trying to do everything.
- 26
LangSmith
Tracing and debugging what actually happened inside a chain or agent run, step by step, instead of guessing from the final output.
Production & Safety
- 27
Evaluation
Measuring whether your agent is actually good, using test sets and metrics, instead of eyeballing a handful of examples.
- 28
Security / guardrails
Protecting against prompt injection, unsafe tool calls, and data leaks — the things that turn a cool demo into a liability.
- 29
MCP
The Model Context Protocol — a standard way to connect models to tools and data sources, so integrations aren't reinvented for every app.
- 30
Production architecture
Deployment, scaling, cost control, and monitoring — what it actually takes to run an agent system reliably in the real world.
Mastery
- 31
Advanced agentic systems
Putting all six themes above together: long-running, self-correcting, multi-agent systems that plan, act, and adapt with minimal supervision.
Where to start
If you're new to this entirely, don't skip to LangGraph because it looks more exciting than "prompt templates." Phases 1–9 are boring on purpose — they're also the reason the agents you build later actually work instead of hallucinating their way through a demo.
That's the whole map: 31 phases, six themes, one system. If you want a deeper breakdown of any single phase, DM me @prashant.code or drop a comment on the reel — whichever phase gets the most questions is probably the next dedicated post here.