Start here
The village is open
Starter Village takes about twenty minutes, needs nothing installed and no account. Or read the whole curriculum first and decide for yourself whether the argument holds.
Someone who wants to build with AI gets handed two separate things: a Python course with no destination, and the news that AI is the thing that matters now. Nobody walks them across the gap between the two.
Agentale is an attempt to build that bridge as a game. This page is the argument underneath it: what we think is broken, what we are betting on instead, the rules we hold ourselves to, and exactly how far the world goes today.
Counted from the game content itself, so this page cannot drift away from what has actually shipped.
The usual route from curious beginner to person who builds with AI asks you to cross a gap that nobody shows you how to cross.
Strings, then loops, then dictionaries, in the order a textbook likes them. Every exercise is correct and none of them is going anywhere in particular. A learner who asks what this is for is told it will make sense later.
At the same time the message everywhere is that AI is what matters now. That arrives as news rather than as something you can act on, and it never once connects back to the print() sitting on your screen.
Between “I can write a for loop” and “I built an agent that decides when to call a tool” is a stretch that almost nobody teaches as one journey. Most people who start a beginner course never reach the end of it, and the ones who do are still standing at the edge of that gap.
Agentale never teaches a concept because it is next in the syllabus. It teaches it because a quest is blocked without it, and because your companion cannot do the thing yet.
You do not complete Python and then begin AI. From the very first quest you are giving abilities to something that will become an agent, and the language shows up as the list of abilities you gave it.
name = 'Wren'How it remembers. A traveller says their name once and your companion still has it three lines later.
if total >= 50:How it decides. Free delivery or not, and it makes the call itself instead of you making it for it.
for order in queue:How it gets through a queue of work without you writing the same instruction twenty times.
def shipping(weight):The skills it keeps. Write one once, and it can be called on every time that job comes up again.
tools=[shipping]How it acts. The model stops describing what should happen and calls the thing that makes it happen.
Only the top row is playable today. The rest is the road, and every idea on it is there because the last row needs it.
These were written before the code and they are what settles arguments about it.
A course with a points system bolted on is still a course. Here the world is the curriculum: you walk into a region, meet someone with a problem, and the lesson is the thing standing between you and helping them.
If learning an idea does not change what your companion can do, that idea has not earned a quest. Abilities are the unit of progress, and each one sits on top of real Python you wrote yourself.
Not a percentage bar. Quests cleared, regions walked, concepts logged in the Dex, abilities your companion now has. Proof of what you can do, in a shape you could show someone.
Early quests hand you a small, honest slice of the real thing and keep the scaffolding out of sight. Later ones lift the cover and show the schema, the loop and the code underneath. Beginner friendly is a promise to come back and explain, never a licence to lie.
Which is why a region ends in a project rather than a quiz, and why you can take its gate test and walk straight past the quests. If you already know it, the gate opens and the badge records how you got through. XP you could not defend is not worth collecting.
Passed the gate test. Complete the quests here and this becomes the Pathwalker badge.
One sentence has settled more decisions here than anything else. It is the transformation the entire world is built to deliver, and every region exists to move someone one step along it.
“I know nothing about coding or AI.”“I understand the building blocks.”“I built an AI agent.”“I can improve, evaluate and eventually ship one.”
“I know nothing about coding or AI.”
Starter Village. You wake a companion and teach it to speak, to remember, to read messy text and to judge a number. Four quests, running in your browser, with nothing to install and no account to make.
“I understand the building blocks.”
Logic Woods is open: decisions, comparisons, loops, lists, dictionaries and sets, six quests deep. So is the Skill Workshop, where your companion learns to build a skill once and use it again, borrow the tools it would rather not build, and read what Python says when something breaks. The pair end in a project where it runs the village ledger on its own, judged on scenarios rather than one expected answer.
“I built an AI agent.”
API Harbor is open: real requests to a harbour we run ourselves, JSON, keys, and a companion that keeps working when the answer never comes back. Prompt Peaks is open: your code talks to a real model, writes prompts and standing rules for it, gets its answers back as data, and checks them before trusting them. Agent District is open too: your code hands the model tools, runs the one it picks, sends the result back, loops until the job is done, and plans before it acts. It ends in the Support Desk, an agent that has to hold up across scenarios, including one that tries to talk it into breaking its own rules.
“I can improve, evaluate and eventually ship one.”
Knowledge Caverns and Production Tower: retrieval, evals, guardrails, cost and tracing. Then a guild, where you build one agent properly and keep it: support, marketing, coding or research.
None of the above is a promise about a date. This is the actual state of the build, in the order it gets made.
Starter Village, Logic Woods, the Skill Workshop, API Harbor, Prompt Peaks and Agent District with its Support Desk boss, with Python running in the browser, a harbour API served from our own routes so nothing costs a key, a real model your code can call, the Dex that logs every concept you discover, the curriculum map, and a gate test you can take to skip a region and still earn the badge.
Knowledge Caverns, where an agent learns to answer from documents it was never trained on.
Everything past the harbour needs a model behind it, and a model costs money per learner. So that was settled before the first AI lesson was written: grading uses a free practice model, the first AI quest is free for everyone, and asking the real model costs a plan’s credits, under a daily ceiling.
Production Tower, then the guilds, which each end in one agent you keep and a title to go with it.
9 lessons past Starter Village are written down and waiting. Writing them properly is the work.
Code is invisible, and that is the actual teaching problem.
A beginner writes a line, a check goes green, and nothing about that tells them what they can now do. Ability stays abstract at exactly the moment it most needs to feel real.
So we put the ability somewhere you can watch it. The thing on screen can do something today that it could not do yesterday, and it can do it because of code you wrote. Not a badge that was handed to you. A capability you gave it.
Four abilities, four Python ideas, in the order the story needs them.
Puts words on the screen with print().
Your companion can greet you and tell you what it is doing.
Stores values in variables and uses them later.
Your companion can remember your name and the things you tell it.
Reads, tidies and reshapes strings.
Your companion can clean up messy input and write tidy replies.
Does the maths and turns comparisons into True or False.
Your companion can work out totals and make yes-or-no calls, like who gets free delivery.
Takes one path when something is true and another when it is not.
Your companion can act differently depending on what it finds.
Compares two values and turns the answer into True or False.
Your companion can say which of two things is bigger, smaller or the same.
Works through a queue with for and while.
Your companion can handle a hundred requests without you writing a hundred lines.
Keeps an ordered pile of things in a list.
Your companion can carry a list of jobs and take them one at a time.
Finds a value by its name in a dictionary.
Your companion can find one traveller's preferences among hundreds.
Keeps one of each, using a set.
Your companion can spot repeats and keep a clean list.
Bundles a set of steps into one named skill.
Your companion can learn a skill once and use it whenever you ask.
Takes what it is handed and gives a result back.
Your companion can do the same job on anything you give it and report the answer.
Brings in code somebody else already wrote.
Your companion can use tools it never had to build.
Reads an error and names the line that broke.
Your companion can tell you what went wrong instead of just stopping.
Sends a request to another computer and waits for the reply.
Your companion can fetch something it was never told.
Reads the data an API sends back and finds the part that matters.
Your companion can pull one number out of a whole cargo of data.
Hands data over to an API instead of only asking for some.
Your companion can file a report, not just read one.
Proves who it is with a key, without putting the key where anyone can read it.
Your companion can use an API that only answers to those it knows.
Meets a failed request calmly and carries on.
Your companion keeps working when the answer never comes back.
Asks a language model a question and reads its answer.
Your companion can ask the oracle anything and pass on what it says.
Builds a clear prompt from parts: what to do, for whom, and how much.
Your companion can ask the oracle exactly the question it means.
Gives a model standing rules and carries a conversation from turn to turn.
Your companion can give the oracle rules to keep and hold a real conversation.
Asks for an answer as JSON and turns it into data, safely.
Your companion can get answers back as data it can act on.
Checks a confident answer against the facts before trusting it.
Your companion can tell a checked answer from a confident guess.
Runs the function a model asks for, with the arguments it chose.
Your companion can let the oracle ask for a tool, and use it on its behalf.
Describes its tools so a model can pick the right one, and follows its choice.
Your companion can offer the oracle several tools and run the one it picks.
Hands a tool’s result back to the model, with everything that came before it.
Your companion can tell the oracle what a tool found, so it can answer in words.
Acts, looks at what happened, and acts again until the goal is done, or stops.
Your companion can work through a job on its own, one tool at a time.
Breaks a big goal into steps before acting, and keeps to a budget while it works.
Your companion can turn a big goal into a plan, then carry it out step by step.
This is a product about building AI agents, made at a moment when it would be easy to ship an AI tutor that hands over the answer. That would quietly defeat the point. A beginner who gets unstuck by a model has not learned the thing they came here to learn.
So the help is written by hand: an explanation before the task, hints that escalate only as far as they need to, Python errors translated into plain language, and a fix step where the broken code is small enough to actually read. Nothing inside a quest writes your solution for you.
Where AI does show up later in the world, it is the subject. You write the prompt, you handle the structured output, you decide when a model is allowed to call your function. It is the thing being studied, not the thing getting you through.
A hint that reads your code and points at the line you got wrong, without writing the next one for you. It is on the roadmap. It is not built, and until it is, it stays marked as such everywhere it appears.
A nudge from an AI tutor that reads your code and points at what to fix.
Starter Village takes about twenty minutes, needs nothing installed and no account. Or read the whole curriculum first and decide for yourself whether the argument holds.