Quoridor Bots — Beginner to Research-Grade

The roster is seven opponents: six difficulty tiers that form a ladder, and one experimental engine that sits off it entirely. The six tiers are the same program — a Monte Carlo tree search, ported from the open-source gorisanson/quoridor-ai project, MIT licensed (how it scores a move) — and they differ in one thing: how many simulated games each is allowed per move, from a couple of hundred at the bottom to 60,000 at the top.

Two things complicate that clean picture, and both are deliberate. The two gentlest tiers are weakened outside their search: a wrapper takes the move the search chose and, some of the time, replaces it with a legible, human-shaped mistake. A small search left alone plays like a strong player in a hurry — consistently mediocre — where a weakened one plays like an actual weaker player, and those are the mistakes worth learning to punish. And from the third rung upward the wrapper is gone and the budget stops being flat: harder positions get more thinking, easier ones less.

Every opponent here is free to play without an account. Each name below links to a full page on how that bot works and how to beat it — and because the ladder is one program at six budgets, each page carries the one lesson that belongs to its rung.

The ladder, bottom to top

Scout — beginner

The gentlest tier, and its weakness is manufactured rather than incidental: roughly a quarter of its moves are replaced with a plain march toward its goal, and another fraction with a deliberate give-back of up to three squares of path. That makes its mistakes the identifiable kind — the kind you can practise spotting while there is still time to punish them. The right first opponent after your first game.

What a plain march throws away. With Red to move, a8h would turn Blue's one remaining step from a8 into three moves; when Scout's march fires it steps to f2 instead, and Blue simply walks in.

Mason — intermediate

The same small search and the same wrapper with every knob dialled down: the largest square it will hand back is one. What defines Mason is its horizon — it answers a threat that already exists and struggles with one still being assembled — which makes it the tier where a quietly sequenced two-wall plan starts to pay. Move here once you can count a race without stopping to think about it.

Warden — advanced

The first rung where nothing is handed back: no weakening touches its move, and its thinking budget flexes with the position. It also declines to search two whole phases of the game — the first few plies come from a small opening lookup, and once its own walls are spent it simply walks a shortest path. Watching where Warden chooses to think is a decent guide to where you should.

Oracle — expert

Three times the playout ceiling of the tier below, and the last full measurement of the ladder was unambiguous about it: Oracle won every game of that match. This is the rung where the interesting question shifts from how much it thinks to where it looks — its candidate walls come from a compact stencil around the pawns, existing walls, and the board edges, a scan worth borrowing for your own play. A lone wall in open ground never enters that stencil at all, which is also the shape missing from its picture of the future.

A wall missing from Oracle's picture of the future. c3h would add a square to Red's walk down the d-file, but it touches no wall, sits in neither pawn's ring and is no outer-column horizontal, so in this position Oracle does not generate it, as its own move or as one you might play.

Magnus — master

Magnus plays the move its search kept returning to, not the one with the flashiest best case, so ideas that need one cooperative reply from it go nowhere. It also triages its own clock — thinking hardest where walls are unspent and the race is close — and that triage rule is worth stealing. One honest caveat from the ladder measurements: the step up from Oracle came out an even split over a full match, so expect Magnus to feel like a sterner version of the same game rather than a wall.

Nemesis — extreme

The strongest opponent on the site, with a 60,000-playout ceiling, and it beat Magnus in every game of the last measurement. It has exactly one documented blind spot: wall pockets that close slowly, where two and a half times the thinking budget failed to improve its score on the site's own trap suite. Very hard to beat by playing well; somewhat easier to beat by playing differently — its page explains what that means in practice.

One of the nine trap-suite positions, from the recorded game against Magnus. Red on e4 is five moves from home, but its own e5h and c5h have shut the way back up, so a single wall, e4v, would make it fifteen; without them the same wall would make it nine.

Off the ladder

Neural (Experimental)

The odd one out: a separate program built around a network that judges positions instead of playing them out — and currently the weakest opponent on the roster, not the strongest. At equal thinking time it loses to every tier it has been tested against, it takes close to eight seconds a move, and it improves in discrete generations rather than continuously. Play it for a winnable first game, or to watch a learning engine be honestly bad at something in public.

Before you sit down

If you have not played at all, start with How to Play — the whole ladder assumes you can count both shortest paths every turn, and that habit is where the guide begins. The strategy series is the preparation the upper rungs quietly demand: wall economy, tempo, and the trap structures that are the one thing even the top of the ladder handles badly.