AI GAME LAB

Real games.
Built by AI agent teams.

The internet is full of "one prompt, one demo" AI games. This lab is the opposite: we run agent teams against production standards — contracts, runtime evidence, honest status boards — and document everything, including what failed.

What is this? The build log of two real games made with AI agent teams — Linjiang Chronicles (a 305-event life simulation) and Moonlit Kingdom (an idle RPG heading to iOS) — plus the pipelines, the economics, and the honest limits of AI game development in 2026.

The games

Linjiang Chronicles narrative choice screen

In developmentLinjiang Chronicles (一纸青云)Ancient-life simulation for iPhone: 305 events, 711 choices, 20 endings. Take the imperial exams, survive flood season, live with the consequences.

Godot 4.6 · Portrait · Premium → Game file

iOS deliveryMoonlit Kingdom (月光王国)Idle RPG: click once, then automate everything. 8 chapters, 40 expedition routes, 16 relics — with published economy formulas.

Godot 4.6 · Landscape · Idle → Game file

Guides & pipelines

Start hereHow we build games with AI agentsThe complete production pipeline: art generation, contract-driven implementation, QA evidence trails — and why demos aren't games.

Flagship guide · 8 min

AI game art that survives the runtimeSprite atlases with frame manifests, chroma-key cleanup, map generation, audio — the pipeline behind every asset in our games.

Pipeline guide · 7 min

The evidence system: QA that keeps AI honestRuntime recordings, same-state captures, source-bound audits — how to verify what an agent claims it built.

Pipeline guide · 7 min

Vibe coding games: what actually works181,000 games shipped, revenue barely moved. The honest split between what AI nails and what still eats projects alive.

Honest guide · 6 min

How this lab works

Three rules, applied to everything we publish:

No claim without runtime evidence. Every feature we call done has a recording of the actual build — not a screenshot from a good day.
Status boards, not hype. Each game page ends with an honest table: what's playable, what's pending, what's not accepted yet. (Our life sim's "20 hours of content" claim is currently flagged as unproven — that's what the boards are for.)
Publish the failures. Cut features, rejected media, automation that wasn't worth it. The failures are where the pipeline knowledge lives.

Why this lab exists

2026's game-dev conversation is dominated by generation: ~20% of Steam Next Fest entries declare generative-AI use, and vibe coding pushed 181,000 mobile games to market in six months. The industry's own post-mortems agree on the gap: making a playable prototype is easy now; making a game that keeps players is still hard. This lab documents the hard part.

What's coming

The QA evidence system — how same-state captures and runtime recordings catch what "looks fine" misses.
The economics — what this pipeline actually costs to run, per feature and per month.
Playable web demos — both games will appear here before anywhere else.

QUICK ANSWERS

Are the games playable yet?

Both run today in development builds. Linjiang Chronicles has complete single-player flows with headless test suites; Moonlit Kingdom is in iOS delivery. Playable web demos will be announced here first.

Were the games fully AI-made?

No — and that honesty is the point of this lab. Agent teams handle generation, implementation, testing, and evidence collection at a scale one person cannot. Humans own taste: what feels fun, what gets cut, what ships. The guides show exactly where that line sits.

Can I use these pipelines for my own game?

Yes — start with the flagship guide, then the art pipeline. The approaches are provider-neutral and scale down to solo projects.