A 22-module path from "what even is an agent?" to a deployed multi-agent system.
Nothing here is a tutorial you watch. Every module is code that has actually been executed against live APIs, with the real output recorded. When something didn't work, the failure is written down too β those failure notes are the most useful part of the repo.
If you are a complete beginner: that's fine. This README tells you exactly where to start and what to ignore. Jump to π― Where to start.
Not Module 1. Here's the honest reasoning:
| Module | What it is | What to do |
|---|---|---|
| 0οΈβ£ Python Core Basics (optional) | Variables, loops, functions, classes, imports | π Only if you're new to Python. Already comfortable? Skip straight to 1. |
| 1οΈβ£ Foundations | Concepts only, no code | π Skim for 20 min. Good vocabulary, nothing to run. |
| 2οΈβ£ Python Essentials | async + Pydantic |
β
Check, don't study. Can you read async def / await, and class X(BaseModel)? If yes, move on. You'll need it from Module 7. |
| 3οΈβ£ LangChain Agents | The old way of building agents | π Skim only. This is the approach LangGraph replaced. Useful as contrast, not as a foundation. |
| 4οΈβ£ LangGraph Fundamentals | State, nodes, edges | π START HERE. Verified working code + beginner notes. Everything after this is built on it. |
Why Module 4 maximises what you already know: you know Python functions and dictionaries. That is genuinely all Module 4 is β a function that takes a dict and returns a dict, plus wiring that says which function runs next. No new mental model required on day one.
Day 1 β Module 4, files 01β02 State, nodes, edges + your first LLM node
Day 2 β Module 4, files 03β06 Chaining, parallel + reducers, routing, loops
Day 3 β Module 6 Memory β the one that makes it feel like ChatGPT
Right now, in your terminal:
cd "Module 4 - LangGraph Fundamentals/project"
uv sync # required: venvs were cleared to reclaim disk
uv run python 01_temperature.pyLook at the output before you open the file. Then read LEARNPLAN.md β it's the method that turns this code into understanding instead of a folder you own.
ββ π’ PHASE 1 Β· FOUNDATIONS βββββββββββββββββββββββββββββββββ
β M1 concepts M2 python M3 old-style agents β skim
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β
ββ π΅ PHASE 2 Β· THE CORE β most important βββββββββββββββββ
β M4 state/nodes/edges β M5 routing & loops β
β M6 memory & threads β M7 tools β M8 RAG + approval β β learn properly
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β
ββ π£ PHASE 3 Β· SHIPPING ββββββββββββββββββββββββββββββββββββ
β M9 Docker/FastAPI M10 capstone M11 multi-agent β
β M12 MCP M13 subgraphs M14 guardrails β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β
ββ π PHASE 4 Β· PRODUCTION CRAFT ββββββββββββββββββββββββββββ
β M15 harness M16 evals M17 gateway β
β M18/20/22 advanced RAG M19 serverless M21 case study β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
π‘ If you only ever do four modules: 4 β 6 β 7 β 8. State, memory, tools, retrieval. Almost everything else in this field is a variation on those four, and Module 10 is literally those four stacked together.
| # | Module | You learn | Status |
|---|---|---|---|
| 0 | Python Core Basics | Variables, types, collections, control flow, functions, a minimal class, imports | β built |
| # | Module | You learn | Status |
|---|---|---|---|
| 1 | Agentic AI Foundations | LLM β RAG β agents β multi-agent; the ReAct loop | π + NOTES |
| 2 | Python Essentials | asyncio, concurrent calls, validated output |
π + NOTES |
| 3 | Agents with LangChain | Tool calling, single + multi-agent (LangChain 1.x) | β verified |
| # | Module | You learn | Status |
|---|---|---|---|
| 4 | LangGraph Fundamentals | State, nodes, edges, reducers | β 6 files |
| 5 | Workflow Patterns | Send fan-out, Command routing, capped loops |
β 5 files |
| 6 | Chatbot: Memory & Streaming | Checkpointers, thread_id, time travel, SQLite |
β 5 files |
| 7 | Observability & Tools | ToolNode, tools_condition, context trimming |
β 4 files |
| 8 | RAG & Human-in-the-Loop | PDF β FAISS, interrupt() + resume |
β 6 files |
| # | Module | You learn | Status |
|---|---|---|---|
| 9 | Deployment | FastAPI streaming, Docker (image verified), CI | β built |
| 10 | Capstone: ChatGPT Agent | All of Phase 2 + Streamlit UI, 4 tools Β· ARCHITECTURE | β built |
| 11 | TripMate AI | Supervisor + 3 specialists, Command(goto=) |
β built |
| 12 | MCP | One tool server, many clients, async, stdio | β built |
| 13 | Subgraphs: AgentWriter | Reusable subgraph + revision loop | β built |
| 14 | Guardrails + Supervisor | 4 middleware layers, 5/5 adversarial eval | β built |
π Phase 4 β Production Craft (15β19 and 21 built; 20 and 22 need Pinecone / Neo4j credentials)
| # | Module | You learn | Status |
|---|---|---|---|
| 15 | Harness & Loop Engineering | Audit 8 rows, fix the weakest in code | β built |
| 16 | Agent Evaluation | 31 evals, 5 dimensions, LLM-as-judge, CI gate | β built |
| 17 | LLM Gateway | Groq + DeepSeek + Gemini, verified fallback | β built |
| 18 | Advanced RAG I | CRAG + Self-RAG, compared side by side | β built |
| 19 | Self-Correcting App | Writer/reviewer loop, container verified | β built |
| 20 | Advanced RAG II | Routing RAG on hosted Pinecone | π plan |
| 21 | Forward Deployed Eng | FDE case study with real before/after | β built |
| 22 | Advanced RAG III | Neo4j GraphRAG + charts/images | π plan |
1. Install uv (replaces pip/venv, much faster):
curl -LsSf https://astral.sh/uv/install.sh | sh2. Get free API keys:
| Key | Where | Needed by |
|---|---|---|
GROQ_API_KEY |
console.groq.com | π΄ 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 19 β free tier |
TAVILY_API_KEY |
tavily.com | 3, 7, 10, 11, 13, 18 β free tier |
LANGSMITH_API_KEY |
smith.langchain.com | 6, 10 β optional, tracing only |
OPENWEATHER_API_KEY |
openweathermap.org | 11 β optional, degrades to search |
PINECONE_API_KEY |
pinecone.io | 20 (not built) |
NEO4J_URI/_USERNAME/_PASSWORD |
neo4j.com/aura | 22 (not built) |
Copy .env.example into each module folder as .env and fill in only what that module needs. Embeddings need no key β Modules 8/10/18 use local MiniLM.
π° Cost: Groq's free tier runs everything here. If you'd rather use OpenAI, a full pass of every module is about $0.09 on
gpt-4o-mini(252 counted model calls) β so the usual $5 minimum top-up is far more than enough. Details and caveats in VERIFICATION.md Β§4.
3. Run any module β each is its own self-contained project:
cd "Module 4 - LangGraph Fundamentals/project"
uv sync # installs exactly the pinned versions
uv run python 01_temperature.pyPut your keys in a .env inside the module folder. .env is gitignored everywhere β never commit it.
β οΈ If a module builds its model client at import, importing it needs a key. That's fine when you run the module β you have a.env. It is not fine for the free CI job, which runs with no secrets at all: it madetests.ymlfail at Module 18 before a single assertion. Module 18 now builds both clients on first use (self_rag.llm(),kb.embeddings()), so the pure-logic tests import cleanly. Reproduce CI before trusting it:env -u GROQ_API_KEY uv run python test_caps.py.
β οΈ Folder names use-, never:β a colon makesuv runfail outright and makessource .venv/bin/activatesilently do nothing (you stay on system Python and get baffling import errors). SeeAGENT_RULES.md#5.
| Doc | Read it when |
|---|---|
| π LEARNPLAN.md | Now. How to actually learn this: the run β break β predict loop, how to read a NOTES.md, realistic pacing, break-it experiments per module, an error decoder. |
| π οΈ AGENT_RULES.md | When building Modules 17β22. Pinned stack + every pitfall already hit. Paste it with a module's README when asking a coding agent to build it. |
| π§ MEMORY.md | Starting a fresh chat? Paste this + AGENT_RULES.md and an agent can continue without re-reading the repo: module status, the pitfall list, remaining work prioritised, and the exact checklist for when an OpenAI key arrives. |
| β VERIFICATION.md | Proof it runs. Real captured output from every built module, what is honestly not verified, the full key matrix, and what it costs (~$0.09 a full pass). Regenerate with ./verify.sh. |
| π OpenAI.md | If you have an OpenAI key. Switching provider is one env var; it also removes the json_mode workaround and the parallel-tool-call limitation. Includes cost warnings and an honest verification table. |
| ποΈ topics.md + learning_prompt.md | Want a chatbot to teach you the theory, not just the code. Paste learning_prompt.md then topics.md into any chatbot for a one-topic-at-a-time walkthrough, prerequisites checked first. |
Every built module also has a NOTES.md written for a beginner β what each file does, why, and a "Gotchas we hit" section at the end. Read the gotchas twice: once now, once when you're stuck.
This repo follows entbappy/Complete-Agentic-AI-Course β the official repo for the course. I scanned it and compared structure directly. The overall idea matches. Same concepts, same order, same projects:
| Original repo | β | Here |
|---|---|---|
Langchain vs LangGraph |
β | M1 |
Asynchronous Programming + Pydantic-Validation |
β | M2 |
Langchain-Single-Agent + LangChain-Multi-Agent-Research-System |
β | M3 |
LangGraph-Code/1,2,3,5,7,8_*.ipynb |
β | M4 (near-exact 1:1 β temperature, QA, chaining, essay, review, iterative) |
LangGraph-Code/4,6_*.ipynb + workflow patterns |
β | M5 |
9_Persistence.ipynb + Agentic-Chatbot/*_db_* |
β | M6 |
Agentic-Chatbot/*_tool_* + tools_demo.ipynb |
β | M7 |
10_HITL.ipynb + *_rag_* |
β | M8 |
Agentic-Chatbot-using-LangGraph/ (full app) |
β | M9 + M10 |
11_subgraphs + 12_subgraph_shared |
β | M13 (the original shows both patterns β M13's notes explain why to prefer the explicit one) |
13_guardrails_crash_course.ipynb |
β | M14 |
AI_Agent_Evaluation.ipynb |
β | M16 |
llm_gateway.ipynb |
β | M17 |
Corrective-Rag-CRAG + Self-Rag-Code |
β | M18 |
Advanced-RAG/Agentic-RAG |
β | M20 |
Agentic AI Interview Preparation/ |
β | M21 |
Advanced-RAG/Graph-RAG + Multimodal-RAG |
β | M22 |
What's different, and why:
- π 22 small modules instead of 12 topic folders. Same material, finer-grained, so one module = one sitting.
- π
.pyscripts instead of.ipynbnotebooks. Runnable with one command, diffable in git, and they fail loudly instead of hiding stale state in cell outputs. - π¦
uv+ pinned versions instead ofrequirements.txt. Reproducible β the reason this code still runs. - β Four modules with no counterpart in the original: M9 (Deployment), M11 (TripMate), M12 (MCP), M15 (Harness), M19 (Serverless). These come from the extended course brief.
- π Secrets are gitignored here. The original has 18 committed artifacts, including 4
.envfiles, pluschatbot.db,__pycache__/andfaiss_db/. Don't copy that pattern β rotate any key you ever push.
Your diagnosis is correct, and it's the main reason this repo re-implements rather than copies. The concepts in the original are sound; the code bit-rots because the ecosystem moves fast. Concretely, what broke while building Modules 4β14:
| π What rotted | What it does now | Fix used here |
|---|---|---|
llama-3.3-70b-versatile |
404 β decommissioned | openai/gpt-oss-120b |
mixtral-8x7b-32768 |
410 β decommissioned | same |
with_structured_output(Model) |
errors: "Tool choice is required, but model did not call a tool" | method="json_mode" + field names in the prompt |
langchain_community.tools.tavily_search |
deprecated | langchain-tavily |
init_chat_model("groq:llama-...") |
dead model behind a live API | live model list, checked |
langchain-community |
prints a sunset warning | still correct for FAISS/PyPDF β no official replacement yet |
The lesson, not the complaint: never trust a hardcoded model name in any tutorial, including this one. Check what's actually live:
curl https://api.groq.com/openai/v1/models -H "Authorization: Bearer $GROQ_API_KEY"All of this is collected in AGENT_RULES.md so you hit each gotcha once, not twice.
- π Check "Gotchas we hit" at the bottom of that module's
NOTES.md. - π Check the error table in
LEARNPLAN.mdβInvalidUpdateError,GraphRecursionError, dead models and friends are all decoded there. - π Run it again. Code bugs fail identically every time; model weirdness varies. That one distinction saves hours.
- π Read the last line of the traceback first, not the first.
Concepts and projects in this repo are based on Boktiar Ahmed Bappy's Agentic AI course, which explains all of this material.
- π» GitHub: github.com/entbappy Β· Complete-Agentic-AI-Course
- πΊ YouTube: @dswithbappy
- π¬ Course playlist: Complete Agentic AI Course