The Learning Path

Joeven is ordered like a degree, compressed. A degree would spend months on each layer. You will spend lessons. The order still matters. Later tracks assume you can write a function, parse JSON, and say what a vector is for.

This path exists so you do not learn RAG as a vendor button, then freeze when ranking looks random. It also exists so you do not learn “multi-agent” as a group chat, then freeze when two roles copy the same file. The academy is a stack. Each layer holds the next.

People confuse the path with a buffet. A buffet says “start anywhere.” That is fine for recipes. It is a bad idea for agents. If you cannot parse JSON, your tool layer is theater. If you cannot say what a probability is, your eval numbers are theater. You may peek ahead. You should not skip the foundation and hope.

The map

Python is not optional. Agents are programs. You will write parsers, retries, tests, and HTTP clients. If you cannot write a function that parses JSON and raises a useful error, you cannot ship an agent. The browser boxes use only the Python standard library. Your laptop will add packages later.

Math is the language of embeddings, attention, loss, sampling, and Bayesian update. An embedding is a list of numbers that stands for a piece of text. You do not need a PhD. You need vectors, matrices, derivatives, probability, and entropy well enough to debug a retrieval system that “feels random.”

Machine learning teaches the habit: data, model, loss, eval. Transformers explain tokens, attention, and why context windows matter. A context window is the maximum number of tokens the model can read at once. LLMs are the APIs, costs, decoding, and failure modes you will live with.

The agent stack, in the order you will meet it:

• Prompting and structured output — how to talk to the model so it talks back in JSON
• Tools and MCP — how the model touches the world. MCP is a way to expose tools over a standard interface. You will meet the idea later, not here.
• RAG and memory — how it knows things that were not in the weights
• Architectures — ReAct, plan-execute, reflection, state machines
• Multi-agent — when to split roles
• Evals and safety — how you know it works and how it fails closed
• Production — tracing, queues, secrets, deploy, cost

Projects are five builds, beginner to advanced. Theory you cannot run is trivia. Treat projects as a portfolio, not as extra homework you skip.

Time

If you already write Python, expect 4–8 weeks at 8 hours per week to finish the core plus two projects. If you are starting from scratch, budget 3–4 months and do every exercise. Those numbers assume you run the boxes and fail sometimes. Skimming quizzes in a weekend is not the same clock.

Sam’s eight weeks

Sam can write Python at work but has not touched probability since school. Week 1 is Getting Started and a Python refresh: functions, JSON, errors, tests. Week 2 is math: vectors and probability until cosine similarity is not a magic word. Cosine similarity is a way to score how close two embeddings are. Week 3 is ML and transformers at a conceptual level — enough to know why a long transcript hurts. Week 4 is LLM APIs, cost, and prompting with a schema.

Weeks 5–6 are tools, RAG, and the agents track: one loop, then costumes. Week 7 is evals and a small production checklist: traces, keys, a spend cap. Week 8 is a project: a tool-using agent with tests. Sam does not start a multi-agent “company of gnomes” in week 1. When RAG ranking looks random in week 6, Sam has math words for it.

If Sam had started at RAG, week 6 would have been “tune the prompt” forever.

Run to execute this in your browser. Nothing is sent to a server.

The first line is 14. The second is the whole map joined with arrows. Then math’s position (2 if you count from 0), RAG’s later position, and True for “math comes before rag.” That boolean is the quiz hiding in a print. Change the list order and the boolean flips. The academy’s order is a choice, not a law of physics — but it is the choice that makes RAG debuggable.

Tip:Use the left sidebar like a textbook table of contents. Skip around only after you have finished Python and the agent loop lessons.

What goes wrong

• Skipping Python because “the model writes code.” The model writes code that you must read, test, and refuse. That is Python skill.
• Skipping math because RAG has an API. When scores look random, the API will not explain cosine or priors.
• Framework first. You learn a brand, not the loop. The brand changes. The loop does not.
• Project first with no evals. A demo on Friday, a silent failure on Monday.
• Multi-agent as a personality test. Split roles only when one loop is honest and overloaded.
• Treating time estimates as guilt. Slow is fine. Skipping checks is not.
• Sidebar tourism. Opening every track, finishing none. The map is ordered so you can stop wandering.

Where this goes next

The next lesson is setup: Python on your machine, virtual environments, API keys, and the habit of treating model output as untrusted. After this Getting Started track, Python is the first long track. Do not wait for tools to learn functions. Do not wait for eval to learn assert. Peek at projects so you know the destination. Come back to the map when you feel lost. The Try it list is the sidebar in text form.

How agents use this

A curriculum is a workflow with a goal. Your goal is a testable agent in the project track. Observations are quiz results and whether a Try it box ran. Actions are “do the next lesson” or “repeat Python.” Stop is “this lesson’s check passed,” not “I opened 12 tabs.”

When you build learning features into an agent (a tutor, an onboarding bot), use the same map discipline.

• Code: store progress as a list of finished slugs, like path above. Do not invent a second map in a prompt.
• Logs: record which lesson, which quiz outcome, which box was run. That is a trace of learning, not a vibe of “user is engaged.”
• Tests: math index is before RAG index. Python is in the list. A “skip to multiagent” button still shows a warning if Python is unfinished — if you build such a button at all.
• Stop conditions: for a student, stop skipping when Python and the agent loop are done. For a tutor agent, stop tutoring and hand off when the user asks for medical or legal advice, or when the budget for tokens hits a cap.

The path is long on purpose. Agents fail in the layers you skipped.