From Prompt to Agent: The Essential LangChain and LangGraph Mental Model

Outline

1. Chapter 1

   The Big Picture: Building Blocks, Chains, and When You Need Them

   Start with a mental model: an LLM app is a kitchen where LangChain supplies the standard ingredients and tools, and LangGraph is the recipe that decides order, repetition, and decisions. You meet the running use case, a customer-support assistant that reads emails and drafts replies. You then learn Concept 1 (models, messages, and prompts) and Concept 2 (chains, which link steps together), along with the assumptions behind them, such as LLMs being stateless and probabilistic.

   Learning objectives and key takeaways

   Learning objectives

   • Describe how LangChain and LangGraph relate to each other and to a raw LLM API
   • Define model, message, prompt template, and chain in plain language
   • Explain why LLMs are stateless and non-deterministic, and what that means for design
   • Build a simple prompt-to-model-to-output chain for the support-email use case
   • Decide when a straight-line chain is enough for a project

   Key takeaways

   • LangChain is the toolbox of standard parts, and LangGraph is the orchestration layer that arranges them
   • Concept 1: a model takes messages in and returns a message out, and prompt templates make those inputs reusable
   • Concept 2: a chain pipes the output of one step into the next, so it fits fixed, predictable workflows
   • LLMs remember nothing between calls, so your application must supply all context each time
   • If the steps are known in advance and never loop, a simple chain is the right tool
2. Chapter 2Free account

   Giving the Model Hands: Tools and the Agent Loop

   A model on its own can only produce text, so this chapter introduces Concept 3: tools, which are functions the model can request to look up data or take actions. You learn how tool calling works step by step, with the model proposing a call and your code actually running it. You then see how repeating this cycle creates an agent, using the support assistant to look up order status and search a help-center knowledge base, and you learn why the loop strains a plain chain.

   Learning objectives and key takeaways

   Learning objectives

   • Define a tool and explain how a model decides to call one
   • Trace the tool-calling cycle from request to execution to final answer
   • Write and attach a simple tool, such as an order lookup, to a model
   • Explain what an agent is and how it differs from a fixed chain
   • Recognize the risks of agents, including cost, unpredictability, and unwanted actions

   Key takeaways

   • Concept 3: tools extend a model beyond text by letting it request actions that your code executes
   • The model never runs the tool itself; it only proposes a call with arguments
   • An agent is a loop of think, act, observe, repeated until the task is done
   • Loops and branching are what a linear chain cannot express cleanly, which motivates LangGraph
   • Use agents when the number of steps cannot be known in advance, and keep the tool set small and safe
3. Chapter 3Free account

   Taking Control with LangGraph: State, Nodes, Edges, and Persistence

   This chapter zooms out to the LangGraph mental model, in which a workflow is a map of stops (nodes), roads (edges), and a shared notebook (state). You learn Concept 4, state, nodes, and edges including conditional routing and loops. You then learn Concept 5, persistence and human-in-the-loop, using checkpoints to remember conversations and pause for approval. The support assistant is assembled into one graph, and the chapter ends with a decision guide for choosing between a plain model call, a chain, an agent, and a graph.

   Learning objectives and key takeaways

   Learning objectives

   • Explain state, nodes, and edges using the map-and-notebook analogy
   • Design a small graph with a conditional edge that routes between steps
   • Describe how checkpointing gives a workflow memory and lets it resume after interruption
   • Add a human approval step before a sensitive action such as sending a reply
   • Choose between a single call, a chain, an agent, and a graph for a new project

   Key takeaways

   • Concept 4: state is the shared data, nodes are functions that update it, and edges decide what runs next
   • Conditional edges and loops give you explicit control that a free-running agent lacks
   • Concept 5: checkpointing saves state so workflows can hold conversations, recover from failures, and pause
   • Human-in-the-loop approval makes high-stakes actions safe to automate
   • Start with the simplest approach and move to a graph only when you need branching, loops, memory, or oversight