Playing Against the AI in Codi Agent: A Desktop Game Collection That Keeps Growing

When an app says it has AI today, it almost always means the same thing: there is a chat box inside it. You type, the model answers, done. That is useful, but it casts AI as an assistant waiting for instructions rather than something genuinely facing you.

Codi Agent uses its AI for one other thing: being your opponent. In the Games menu, Codi is not an assistant helping you win. It is the player on the other side of the board, reading how you play and trying to beat you. There are six games today, and the list grows from release to release.

Quick summary

• The Games menu in Codi Agent holds six games: Tic-Tac-Toe, Connect 4, Codi Maze, Codi Box Duel, Rock-Paper-Scissors, and the Codi Reasoning Duel.
• Your opponent is Codi itself, not another human. Each game has its own level, win/loss/draw record, and progress saved locally.
• Codi’s moves are computed by a deterministic local engine, so a game never waits on the internet, a database, or a model’s answer.
• The language model acts as a background strategy coach, not the thing that picks moves. When the model is slow, switched off, or offline, the last valid profile keeps being used.
• The collection grows with each release: Codi Maze in 1.1.33, the Reasoning Duel in 1.1.37, Codi Box Duel in 1.1.39, and Rock-Paper-Scissors in 1.1.40. Chess is being prepared next.
• A practice mode is available with no risk at all; coin wagers are optional and still run on a local, beta coin economy.
• All of it is part of Codi Agent, free for Windows 10/11 (64-bit), downloaded at agentdl.codiosity.com.

The six games you can play today

Each game stands on its own: its own level, its own statistics, and a Codi playing style trained specifically for that game. You do not have to finish one to unlock another.

1. Tic-Tac-Toe with a three-mark rule

The goal is still three in a row, but one change rewrites the whole game: each player may only have three marks on the board. When you place a fourth, your oldest mark is removed, and Codi follows the same rule. The board never fills up, there are no draws from running out of squares, and you have to think about which of your marks is about to vanish. The mark order is shown on screen so you can plan it instead of memorising it.

2. Connect 4

Drop a disc into the column you choose and line up four horizontally, vertically, or diagonally. The classic, and the clearest benchmark for feeling how Codi’s style shifts as its level rises: from simply blocking you to setting up double threats.

3. Codi Maze

The only real-time game on the list. You chase bytes through a neon maze, dodge patrolling bugs, and look for a safe route before your energy runs out. Controls accept WASD, arrow keys, swipes, or an on-screen D-pad. A round waits for your first direction before it starts, play pauses when the window loses focus, and checkpoints are saved periodically so a round can be resumed after you close the app.

4. Codi Box Duel

A lines-and-boxes game: you pick one line per turn, and whoever closes a box scores a point and gets a bonus turn. This is a game about chains, because handing Codi one free box often means handing over five. Board size can be set to 3x3, 4x4, or 5x5, and your choice is remembered for the next match.

5. Rock, Paper, Scissors

A best-of-3 duel that looks trivial but shows Codi’s pattern-reading side most clearly. Codi locks its choice first, computed locally and deterministically from the history of previous rounds, and only then is your choice revealed. Your pick in the current round is never used to decide Codi’s move, so Codi is not peeking at the button you just pressed. If you have a habit, this is where it shows.

6. Codi Reasoning Duel

Not a board game but a duel of puzzles: patterns, deduction, focus, and memory, in sessions of 3 or 5 rounds. Answer keys are checked by the local engine, while the language model is used to design how Codi plays, and the challenge level follows your progress. There is a practice mode if you only want the exercise without thinking about levels.

How Codi plays: two brains with different jobs

Every game card has an expandable "How Codi plays" section. What is in there matters, because it holds the engineering decision that makes this feel different from the usual "game with AI": the moves and the model are kept separate.

The first brain is the local engine. It is what actually decides Codi’s move: evaluating the position, searching paths, and guaranteeing every move is legal. It all runs on your device and is deterministic, so play never hangs waiting for an answer from the internet.

The second brain is the language model, and its job is not to play. It works in the background as a strategy coach: summarising the matches that already happened into a validated numeric profile, meaning a playing style (balanced, aggressive, defensive, control, trickster, or adaptive) plus weights for offence, defence, control, traps, and prediction. That profile is what the local engine reads when it chooses a move.

One small detail keeps play fair: once a round starts, the strategy profile it uses is frozen until that round is settled. The coach may refresh the profile in the background, but the refresh does not walk into a match already in progress. Your opponent does not quietly change on move five.

Why the games keep working without internet

Because of that split, the answer is simple: only the coaching needs a model, not the playing. If the model is slow, fails to answer, is switched off in settings, or you are offline, the local engine keeps using the last valid profile and the game runs as usual. If there is no profile yet, a built-in default is ready to play against while the strategy memory starts building.

The model provider in Codi Agent is your choice too, including choosing not to use a model at all. The Games menu is never a hidden reason to turn on an outbound connection.

Levels, practice mode, and coin wagers

Each game keeps its own level, and that level only moves one way: winning raises Codi’s level for the next round, while a loss or a draw never lowers the challenge. So your level is a record of how far you have come, not a number that drifts with your mood.

Every 10 levels, two things happen at once. The reward multiplier steps up with no artificial ceiling: levels 1 to 9 pay double, 10 to 19 triple, 20 to 29 quadruple, and so on. At the same time, one new strategy pattern opens up for Codi, so a high-level opponent is not merely "more careful" but has more ways to play.

Before playing, you pick practice mode or set your own coin wager. Practice carries no risk: results still count towards statistics and still feed the AI’s learning, but your level and win streak stay put. If you do wager, the screen shows your balance, the multiplier, total reward, net gain, what a loss costs, and the draw-refund rule before the round begins, so nothing is a surprise afterwards.

Progress that survives closing the app

Every game saves its session locally and can be resumed. Close Codi Agent mid-match and the game card marks it unfinished and offers "Resume". Codi Maze saves checkpoints periodically and when you leave the page, and final results are guarded so they can never be recorded twice.

Even a small thing like the opening draw is stored with the session. Tic-Tac-Toe, Connect 4, and Codi Box Duel randomise who moves first in every new match, and that draw does not change just because the app was closed and reopened.

What Codi learns, and what never leaves the app

The strategy coach only sees compact game telemetry: round outcomes, statistics, and playing patterns expressed as numbers. Chat content, to-dos, notes, raw personal data, raw boards, and full file paths never enter the coach’s prompt. That boundary is enforced on the app side rather than left to the model’s good behaviour.

Beyond that, learning can be paused from Settings, and the activity that has been recorded can be reviewed in the activity memory timeline and then deleted. So if you would rather play without Codi learning anything from how you play, that option exists.

The "keeps growing" part is a release history, not a promise

The Games menu did not arrive all at once. It grew piece by piece over recent weeks, and the history is visible in the release notes inside the app:

• Version 1.1.33: Codi Maze arrives as an original maze game, and introduces the split between the local engine and the strategy coach.
• Version 1.1.34: the three-mark rule in Tic-Tac-Toe is enforced properly and Codi Maze movement is smoothed out.
• Version 1.1.35: practice mode, coin wagers, and a new strategy pattern every 10 levels.
• Version 1.1.37: the Codi Reasoning Duel lands with pattern, logic, focus, and memory rounds.
• Version 1.1.39: Codi Box Duel is added as pattern and decision practice.
• Version 1.1.40: Rock, Paper, Scissors arrives as a best-of-3 match.
• Version 1.1.41: every game gets a calmer interface, the opening draw is stored with the session, and Codi Box Duel gains 3x3 to 5x5 board sizes.

Chess is next, and the app states its status plainly: chess needs proper move validation and a serious engine, the progress foundation is ready, but it is not part of this version. We would rather write it that way than pin up a "coming soon" card that never lands.

Limits worth knowing before you try it

• Your opponent is only Codi. There is no human-versus-human mode, online or on the same device.
• There is no global leaderboard. Because progress is stored locally, your statistics are not compared with other users and do not travel to another device.
• The Reasoning Duel is thinking practice, not a clinical IQ test and not a guarantee of IQ improvement.
• The coin economy is still local beta; paid coins are not enabled.
• Codi Agent is currently Windows 10/11 (64-bit) only; there is no macOS or Linux build yet.
• The releases carrying this feature are still marked beta in their release notes, so you are genuinely trying something under active development.

Why a to-do app needs games at all

A fair question. The first reason is simple and human: Codi Agent is built to keep you company through a working day, and a five-minute break that does not drag you into social media has real value. A game of Connect 4 finishes before your focus does.

The second reason is more technical, and to us it is the point. Games are the best way to show how we use AI: the model is used where being wrong is safe, while the things that must be right (move legality, match results, coin arithmetic) stay with a deterministic local engine. It is the same pattern we use when building AI features for clients, except here you can inspect it yourself without a demo call.

How to try it

1. 01Download Codi Agent at https://agentdl.codiosity.com/ for Windows 10/11 (64-bit).
2. 02Run the installer. If a "Windows protected your PC" SmartScreen prompt appears, choose "More info" then "Run anyway". That is about the installer’s certificate, not the file itself.
3. 03Open Codi Agent, then pick the Games menu in the sidebar.
4. 04Choose a game, open the "How Codi plays" section to see the state of its AI, and start with practice mode first.
5. 05Once installed, later updates are offered from inside the app, including when new games are added.

Closing

The games in Codi Agent are not an AI gimmick bolted onto a to-do app. They are an opponent you can genuinely play against, built on an honest division of labour between a local engine and a model, and the collection really does grow, something you can verify in the release history rather than in a marketing promise.

And if you are weighing up AI features for your own product, the pattern here transfers: the model for the flexible parts, a deterministic engine for the parts that must be correct. The first conversation is free, with no obligation.

• Download Codi AgentThe official download page for Windows 10/11 (64-bit). Free, Games menu included.https://agentdl.codiosity.com/
• Start with what Codi Agent isWhat Codi Agent is, what local-first means, and how to install it on Windows.https://codiosity.com/en/about/insights/codi-agent-aplikasi-to-do-desktop-gratis-lokal-first
• Services & PricingIf you need AI features or an app of your own: the service list with openly shown prices.https://codiosity.com/en/services
• Contact CodiosityA short, free chat, including questions about Codi Agent.https://codiosity.com/en/contact