Welcome to Joeven

Joeven is a guided academy for autonomous AI agents. An autonomous AI agent is software that uses a language model as a brain, tools as hands, and a loop as a heartbeat. A language model (often called an LLM) is a program that predicts the next piece of text. A tool is a function the program is allowed to call, such as “look up a ticket” or “run a test.” A loop is the repeat of: look at the world, decide, act, look again.

This site exists because a chat box is not an agent. A chat box answers when you type. An agent can keep working while you are not typing: it reads a log, calls an API, checks a result, and only then writes a reply. Teams that skip that difference ship demos that cannot stop, cannot be tested, and cannot be billed honestly. Joeven starts at that difference and stays there.

People confuse Joeven with three other things. It is not a prompt-meme blog. It is not a CSS course. It is not vendor documentation for one agent framework. You will use Python, a little math, and a lot of boring checks. The cinematic part — the model “thinking” — is the smallest piece of a real agent. The rest is code, logs, tests, and stop rules.

How a Joeven page works

Every lesson is one idea. You read it. You run the live Python box in the browser. You answer a short check. You continue when you are ready.

The live box runs Pyodide. Pyodide is Python inside the browser. There is no server and no API key. Output appears only if the code calls print. Your progress is stored in this browser. Complete track quizzes at 80% or more to unlock a certificate you can download.

Read the table once more in words. The teaching text gives you a name. The table shows the moving parts. The walkthrough is a story with a goal. The Try it box is a tiny program. The quiz is a stop condition for the page. That is the same shape you will later put around a model.

The stack you will master is the stack that actually ships agents:

1. Python — the language of tools, APIs, tests, and orchestration. Orchestration means code that calls other code in order, with retries and limits.
2. Mathematics — vectors, probability, gradients, information. You need enough to debug retrieval that “feels random.”
3. Machine learning and transformers — how models represent text and generate the next token. A token is a chunk of text the model reads or writes.
4. LLMs, prompting, tools, RAG, memory — the agent toolkit. RAG means retrieval-augmented generation: fetch facts, then generate.
5. Architectures — ReAct, planning, reflection, multi-agent systems. These are costumes on the same loop.
6. Production — evals, safety, cost, tracing, deploy. An eval is a test for model behavior. A trace is a log of every thought and tool result.

A first evening on Joeven

Walk through a real student, not a slogan. Maya wants a support agent that can look up invoices. She has used ChatGPT. She thinks “agent” means a nicer chat box with a logo.

She opens this page. She learns that an agent is a loop with tools, not a skin on a chatbot. She runs the box below. It prints the track list and walks it like a tiny observe/act loop: see a name, mark it seen, go to the next name, stop at the end. That is not a model. It is the shape. She answers the quiz. The browser stores that she finished Welcome.

She does not paste an API key. She does not skip to multi-agent. She opens the next lesson, which defines an agent with four pieces: goal, observations, actions, and a policy. If she is rusty at Python, she still finishes this Getting Started track first, then does Python before math. That order is the product.

The money-honest version of this site is the same as tutorial sites for twenty years:

• Free tutorials (this is the product that ranks and teaches)
• Sponsor slots and a job board for agent-engineer roles
• Certificates that signal you actually finished the work

You can learn everything here for free. The certificate is proof of completion, not a job offer. Treat the projects as a portfolio.

What you will be able to build

By the last project you will have built:

• A tool-using ReAct agent. ReAct means the model emits a thought and an action, then sees the result.
• A retrieval (RAG) support agent with citations
• A multi-agent software team (planner, coder, reviewer)
• An ops agent with human approval gates
• A long-running personal agent with memory

Those builds sit at the end. The path to them is not a weekend of copying YAML. It is functions that parse JSON, budgets that stop loops, and tests that fail when the agent lies.

Tip:Do not skip Python or math if you are rusty. Agents fail in boring ways — JSON, off-by-one chunking, bad probabilities — not in cinematic ways.

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

The first line of output is the count: 14 tracks. Then the loop prints each name after the word observe, and a reminder to open that track later. It ends with a stop line and a next action. Change a name and run again. The loop does not care what the strings are. That is the point of a list and a stop.

What goes wrong

• Skipping the Try it box. Reading is not running. If you never press Run, you will treat slogans as skills.
• Jumping to multi-agent first. A swarm of confused loops is still confused. Learn one loop with two tools.
• Pasting API keys into Joeven. These boxes are a classroom. Keys belong on your machine, in environment variables, later.
• Expecting a GPU. You do not need one for this academy. You need print, JSON, and a stop condition.
• Treating the certificate as the goal. The goal is a program you can test. The certificate is a receipt.
• Copying quiz answers. The quiz is a stop check. If you guess, the next lesson will feel random.
• Confusing Joeven with a framework tutorial. Frameworks come after you can write the loop in plain Python.
• Stopping after Welcome. Welcome is a map. The definition of an agent is the next page.

Where this goes next

The rest of Getting Started defines the agent, the loop, the difference from chatbots and workflows, the learning map, setup and secrets, the triad of tokens/tools/goals, and a tiny agent with a fake model. After that, the Python track is not optional. Math and ML come before RAG so retrieval is not a black box. Tools, eval, and production are where shipped agents live. This page does not teach those tracks. It names them so you do not wander.

How agents use this

Joeven itself is an agent-shaped product. Your goal is “I can ship a small, testable agent.” Each page is an observation. Running the box is an action. The quiz is a stop check for that page. Progress in the browser is memory. If you skip checks, you drift, the same way a model drifts when nobody verifies the goal.

When you write agent code later, copy this page’s habits, not its marketing.

• Code: one idea per function. Names like goal, step, budget, trace. No mystery helpers named data or tmp.
• Logs: print what you observed and what you did. If a run cannot be replayed from logs, you cannot improve it.
• Tests: a function that returns true when the goal is met. Write it before the loop. Welcome’s quiz is that idea in English.
• Stop conditions: max pages you skip, max hours you spend without a working print, no keys in the browser. Later: max steps, max cost, handoff to a human.

A student who finishes this page and immediately opens a 40-tool framework has not started. A student who can explain the table above, run the track-list loop, and say what Joeven is teaching has started.