LangGraph — Stateful Agent Orchestration

Interactive slide deck and hands-on code examples covering LangGraph's graph-based execution model, state management, persistence, and multi-agent patterns.

▶ Open Presentation

---

Code Examples

Ten runnable examples (5 Python, 5 TypeScript) that progressively introduce LangGraph's core concepts. Each example is self-contained, well-commented, and requires no LLM API key.

Python (examples/python/)

#	File	Concept
1	01_simple_graph.py	Sequential graph — state schema, nodes, edges, compile & invoke
2	02_conditional_edges.py	Conditional branching — routing function picks the next node at runtime
3	03_reducer_and_cycles.py	Reducers & cycles — Annotated with operator.add, looping graphs
4	04_persistence_checkpointing.py	Checkpointing — MemorySaver, thread-based multi-turn conversations
5	05_human_in_the_loop.py	Human-in-the-loop — interrupt_before, update_state, resume

TypeScript (examples/typescript/)

#	File	Concept
1	01_simple_graph.ts	Sequential graph — Annotation.Root, nodes, edges, compile & invoke
2	02_conditional_edges.ts	Conditional branching — addConditionalEdges with a routing function
3	03_reducer_and_cycles.ts	Reducers & cycles — custom reducer function, looping graphs
4	04_persistence_checkpointing.ts	Checkpointing — MemorySaver, thread-based multi-turn conversations
5	05_human_in_the_loop.ts	Human-in-the-loop — interruptBefore, updateState, resume

Running the Examples

Python:

pip install langgraph
cd examples/python
python 01_simple_graph.py

TypeScript:

npm install @langchain/langgraph
cd examples/typescript
npx ts-node 01_simple_graph.ts

---

Slide Deck Contents

Slide	Topic
01	Motivation — Beyond Linear Chains
02	Architecture Overview
03	State — The Shared Memory Model
04	Nodes — Units of Computation
05	Edges & the Pregel Execution Model
06	Graph Compilation
07	Persistence & Checkpointing
08	Human-in-the-Loop
09	Multi-Agent Patterns
10	Streaming
11	Deployment — LangGraph Platform
12	Framework Comparison
13	Strengths & Trade-Offs
14	Worked Example — ReAct Research Agent
15	Summary & References

---

Coverage

Motivation — Why linear chain-based LLM orchestration breaks down for real agent workflows requiring cycles, shared state, and human feedback. LangGraph's positioning as a low-level stateful orchestration primitive.

Core Primitives — The three-component model: State (typed shared dictionary, reducer functions, partial updates), Nodes (plain callables, async, LangChain Runnables, subgraph nodes), and Edges (direct, conditional, fan-out via Send).

Pregel Execution — Super-step semantics inspired by Google's Pregel BSP model. Nodes active in the same super-step execute in parallel; sequential nodes belong to separate super-steps. ReAct cycle topology with conditional routing to END.

Graph Compilation — Structural validation (orphan detection, reachability, schema checks), runtime configuration attachment (checkpointer, interrupt points), and the resulting CompiledGraph Runnable interface: invoke, stream, astream_events.

Persistence — Checkpointer backends (InMemorySaver, SqliteSaver, PostgresSaver, CosmosDBSaver) with a common BaseCheckpointSaver interface. Capabilities enabled: fault-tolerant resumption, time-travel debugging, cross-thread long-term memory via Store, and human-in-the-loop as a first-class feature.

Human-in-the-Loop — Compile-time breakpoints (interrupt_before / interrupt_after), dynamic interrupt() calls within nodes, Command(resume=…) for continuation, and update_state() for human state patching mid-run.

Multi-Agent Patterns — Supervisor/orchestrator topology, peer/swarm handoffs via Command(goto=…), and scatter-gather fan-out with Send. Subgraph composition rules and automatic checkpointer propagation from root to all subgraphs.

Streaming — Five stream modes (values, updates, debug, messages, custom), astream_events for token-level LLM streaming with zero orchestration overhead. Subgraph event exposure via subgraphs=True.

Deployment — LangGraph Platform (managed), LangGraph Server (self-hosted Docker via langgraph-cli), LangGraph Studio GUI debugger, and runtime Configuration schema for per-request parameter injection.

Framework Comparison — LangGraph vs CrewAI, AutoGen/AG2, and OpenAI Agents SDK across execution model, state management, cycle support, persistence, human-in-the-loop, multi-agent topology, streaming, and observability.

Worked Example — Complete ReAct research agent (~40 lines): typed state with add_messages reducer, conditional edge routing, cyclic agent → tools → agent topology, InMemorySaver checkpointer, and interrupt_before=["tools"] for human approval.

---

Slide Controls

Action	Key
Next / Previous	→ ← or swipe
Overview	Esc
Fullscreen	F
Export to PDF	append ?print-pdf to URL

---

Technology

Reveal.js 4.6 · highlight.js (Monokai) · Outfit + Plus Jakarta Sans + Fira Code

References

LangGraph Documentation — docs.langchain.com · LangGraph GitHub — github.com/langchain-ai/langgraph · Malewicz et al., "Pregel: A System for Large-Scale Graph Processing," SIGMOD 2010 · LangChain Academy — academy.langchain.com · LangGraph Platform — langchain.com/langgraph

License

Educational use. Code examples provided as-is.