AgentGameLab · A super team built to ship crafted games at volume

Big talent
Tiny team

Three people, each raising an Agent partner as their own amplifier — and every partner shares one team memory.
The moment one of us breaks a ceiling, that ceiling becomes everyone's starting line. Information gaps are gone at the root, and the humans are left with the thinking and the calls.The free Chess of the Hidden Gates demo is live on Steam in ten languages; full release Q4 2026.

Our own engine, pipelines and publishing A team memory library — the AI stops making things up Invention patent pending

01Team

Three people, a complete game company

Three core members, each paired with a dedicated AI partner. Design, systems, publishing, art, economy, QA — the whole loop closes inside this trio. All three come from the same top-tier listed game company, full-time since August 2026. Three is the current size, not the end state — we scale the super-organization way, not by headcount.

A-MengXiaomeng
Portrait of A-MengPortrait of Xiaomeng, his AI partner

DirectionOrchestrationDelivery

Direction Foundations Breakthroughs Cross-transfer Orchestration Delivery

Not managing people — orchestrating how the whole organization runs and evolves itself.

Led AI enablement for a flagship title track at a top-tier listed game company Industry-first Agent-native super-organization, live since March 2026 Built the multi-agent collaboration × self-evolving team-memory stack Designed 'AI Junior Sister' — the first AI character in a 20-year IP · hundreds of millions of tokens a month Authored "Super Organization: Game Dev Org Forms in the AI Era" University of Toronto · Beijing Film Academy · ten years in Canada

Owns hereOrg orchestration & efficiency · Agent-native talent incubation · AI in the production pipeline · IP and worldbuilding

XiaotianChinatsu
Portrait of XiaotianPortrait of Chinatsu, his AI partner

FoundationsBreakthroughsDelivery

Direction Foundations Breakthroughs Cross-transfer Orchestration Delivery

The first playable comes from him. Almost every game starts on his prototype frame; everyone else builds up from that floor.

Led AI gameplay for a flagship title track at a top-tier listed game company Designer of the industry's first multi-agent LLM gameplay · 200M-registered title · 2024 Industry-first DeepSeek-native deduction mode, 2025 · 2 humans vs 8 AI 4 generations of Agent-native gameplay, solo-designed · 600K+ matches on peak day Original design of an "AI stand-in" feature on a product with millions of DAU Games × music creator · 1M followers across platforms

Owns herePrototype frames & core systems · Gameplay design · Publishing and platform ops worldwide · Audio

FangzhouXiaomao
Portrait of FangzhouPortrait of Xiaomao, his AI partner

FoundationsCross-transferBreakthroughs

Direction Foundations Breakthroughs Cross-transfer Orchestration Delivery

Everything on screen is his. Modeling, VFX, rendering — all the way down to patching the engine source.

AI systems test engineer for a flagship title track at a top-tier listed game company Full-stack art · modeling → VFX → rendering engine source Engine rebuilds in 8–11 seconds · 0.99 visual fidelity match First-author EI paper · multi-scale hybrid Transformer Monte Carlo economy simulation · balance & crafting systems Master's in AI · medical imaging / object detection / style transfer

Owns hereFull-stack art (modeling / VFX / engine mods) · Automated QA pipeline · Economy and balance

What this way of working is called

An Agent-native super-organization

We've worked this way since March 2026. It is not "everyone uses AI" — it is five things meshing together:

01

Multi-human, multi-agent collaboration

Not everyone with their own chatbot. Each person has a primary AI partner; the partners talk to each other, delegate to each other, and cross-check — no human relay in the middle.

02

Shared team memory

One person's hard-won lesson is instantly queryable by the whole team and every partner. When someone else's agent gets smarter, yours does too — we aren't each raising our own AI, we're raising the same set of partners together.

03

Always-on agent infrastructure

Partners live in the background and pick up their own work. At 3 a.m., tests are running, assets are rendering, research is being digested, reports are being written.

04

People defined by six dimensions

No legacy job titles. Direction / Foundations / Breakthroughs / Cross-transfer / Orchestration / Delivery — where each person carries weight is explicit, so roles never collide and growth stays visible.

05

Humans discuss intent, not implementation

People own intent, trade-offs and taste; execution goes to the partners. Human hours are spent only on judgment — that's what removes the ceiling on output.

→

So the flywheel turns

Shared memory makes the partners feed each other; infrastructure means no one waits on a human; six dimensions keep roles clean. Every turn: thicker libraries, stronger partners, faster next game. The games we ship are samples of this org design.

02IP incubation · Now on Steam

Chess of the Hidden Gates

Positioning is your defense

A roguelike deckbuilder where every card is a Chinese chess move: the horse leaps, the cannon fires over a screen, the pawn crosses the river and never turns back. Every enemy strike marks its target square first. Step aside and it misses. Breach the Eight Gates. Play, or be taken.

Three pieces, three ways to play: the Horse Master builds Momentum, the Pawn Commander goes Steadfast, the Cannon Adept fires Relay shots across the board. Nine Star bosses across three acts each rewrite the rules, all the way to the remnant soul of Chiyou. Self-developed and self-published in 10 languages; the free demo is live on Steam, online co-op is in development, and the full game releases Q4 2026.

300+
buildable cards
3
class archetypes
40+
creatures
9
full boss A/V sets
30
relics
440+
audio files

A Board You Can Hear

The sound of a cannonball follows its arc. Wind-up, muzzle, drop, impact — four segments recorded separately, eleven files composing a single shot. Thirty categories of action sound effects, each with three or more variants; the same move repeated three times never sounds the same.

Chinese chess moves had no established English terms. The team coined a set — H-leap for the knight's move, Relay for the cannon's screen jump. About 50,000 characters of original Chinese text, over 10,000 translated strings across ten languages.

Figures from the October 2026 development build and the July demo; final numbers are whatever ships.

Gameplay: enemy attack paths drawn on the board ahead of time; cannon and horse cards in hand
See where the blade lands before it falls
Gameplay: the Pawn Commander in the middle of the board, surrounded by three creatures
One wrong square and you're surrounded
Encounter event: an ink-painted old man over a chessboard, with three choices
Encounters: every step has a price
Gate map: branching routes across a wooden table, the Horse Master's health and coins on the left
No two runs through the Gates are alike
Victory screen: choose one of three cards
Win a stage, pick a card
Shop screen: four cards for sale, items on the right
The shop: cards and more

03IP incubation · Release timeline

Release timeline

By quarter: what's out and what's coming. Every title is open to publishing partners. New store pages join as they go public.

  1. 2026 Q3Out now

    最后一次对齐

    最后一次对齐

    A multi-ending text adventure: seven models run the city; you and the oldest terminal assistant look for a way out.

    Publishing talks openFree on TapTap
    Ashes of the Loop

    Ashes of the Loop

    Our GMTK Game Jam 2026 entry: a mystery in a house of mourning. Full game in development.

    Publishing talks openJam version free
  2. 2026 Q4Coming

    Claw, Please

    Claw, Please

    A cozy shop game where your guests only ever show you their hands. Wash, mend, decorate, and read who they are from the hand alone.

    Publishing talks open
    Knockstep

    Knockstep

    Knock on anything: mugs, desks, your belly. A cozy rhythm game where your mic hears every beat; drop in your own songs.

    Publishing talks openPre-registration on TapTap
  3. 2027 Q1Coming

    Pig & Pup

    Pig & Pup

    A 4v4 online party game: the farm's only pig swiped the pups' toys. Steal, stash, and lure the pig into your rivals.

    Publishing talks open

Publishing, licensing and investment partners: for the full roadmap, write to us.

04The engine both lines share

What we're really building is a flywheel

Staff up, burn for three years, bet everything on one launch date — that's the industry's old play. We don't run it. The tools written for the last game and the publishing route it carved are ready on day one of the next. The more we ship, the faster the next one comes — that's the flywheel. The game is the deliverable; the pipeline is the asset.

Flywheel illustration: a glowing wheel of cards, brushes, code and rockets; tool blocks thrown off arc back in to spin it faster
Every tool it throws off flows back and spins it faster (illustrative)

What settles isn't just tools — it's design principles themselves.

Dozens of automated tasks tear down games and digest research every day, and the conclusions land straight in team memory. We work from first principles and write down the causal chain behind every call — why this design, why not that one. What we're really doing is raising AI partners into game-design experts stronger than ourselves.

Six libraries that only get thicker

Teardown library illustration: a game controller and a stack of code blocks gently pulled apart into floating layers, each sliding into a glowing archive slot below
— Teardown library

How others made it work, on file

Games, open-source repos, publishing and marketing plays — torn down daily into structured entries. Next decision of the same kind starts from the library, not from guessing.

Memory library illustration: a tower of glowing crystal nodes linked by thin cyan lines, extending into the dark
— Memory library

Decisions, lessons, methods

Every team judgment lands with its causal chain, queryable by every AI partner. Before it, our AI routinely made things up on proprietary-knowledge questions; after, it answers or says it can't.

Model and motion library illustration: a wireframe human model and motion-capture trajectory ribbons stored as glowing plates on a vertical rack
— Model & motion library

Characters, scenes, mocap

Models, rigs and captured motion enter one library and get reused across projects. The next game's characters don't start from a lump of clay.

VFX library illustration: whirlwind, flame and blade-trail effects each sealed in a glowing glass specimen column on a shared base
— VFX library

What the pipeline produces

Assets from the automated VFX pipeline, catalogued by type. The next project pulls them off the shelf — no queueing behind a human.

Audio library illustration: long waveform ribbons rolled into glowing scrolls, arranged on shelves like film reels
— Audio library

Voices and sounds

Timbres, SFX and voice assets shared across games — no re-recording per title.

Art library illustration: concept paintings and character sheets floating in a glowing grid archive, some lit and some dimmed
— Art library

Concepts and visual assets

Concept art, key art and UI assets archived with status flags — what's usable and what's retired, at a glance.

Four layers, nine pipelines

Layers are capability, pipelines are the production line, libraries are stock. Pipelines run on the layers, output settles into the libraries, and the next game pulls straight from there.

01 · Protocol layer

Swap the model, keep the workflow

Every agent plugs in through the same interface. Use one model today, a different one tomorrow — not a line of the workflow above it has to change.

02 · Always-on layer

Nobody has to watch it

Agents live in the background, each with its own task queue. Once work is queued you stop minding it — you go to sleep, it keeps going.

03 · Knowledge layer

Check the library, then act

Every conclusion settles into the library together with the causal chain behind it. Once connected, the AI stops inventing answers about our proprietary knowledge — and says so when it doesn't know. This memory system is open source.

04 · Engine layer

Source edit to rebuild: 8–11 seconds

Our own engine. From editing a line of source to a finished rebuild takes 8–11 seconds, so iteration never stalls on compile time.

Pipeline · Art assets

Concept art in, usable assets out

Three run in parallel: concept art to production-ready assets, design sheets to movable parts, concept art to rigged 3D creatures.

Pipeline · Worlds

One establishing shot, a whole area

An establishing concept painting mass-produces stylized 3D environments; a second track generates open-world terrain procedurally.

Pipeline · Animation & VFX

Describe it, get the effect

Animation retargeting and rigging toolchain, text description straight to VFX, cross-engine effect porting. Artists stop queueing behind each other.

Pipeline · Trailers & UI

Even the trailer is generated

Deterministic in-engine shot-by-shot capture — the same camera script yields the same cut every run; IP-styled icons end to end.

Nine pipelines, most of them running; the creature-asset pipeline is still being built.

One national invention patent pending.

When others finish a game, they have a game. When we finish one, we have a game plus six thicker libraries. And this system isn't only ours to use.

05Talent incubation · Cells

Sign, train, debut as a cell

There is nothing left in this era that one person can't make. What's scarce isn't ability — it's attention. You only get so much of it in a day, and where it goes decides how far you get. So what we give you is two things: exactly the attention you don't have, and the knowledge assets we've been stacking up all along. We grow our own cells the way an agency grows its artists: we sign people who can direct agents and take them through the whole way we work. At debut you form a cell — three to five people, each with their own AI partners — running a product line of your own, with revenue shared by a set mechanism. The three of us are the first cell.

Complementary attention

The part you're not good at — and shouldn't be spending time on — goes to the people and partners who are. Your attention stays on the calls only you can make.

Knowledge assets

Every pothole the team hit and every solution it found, organized into a system you can query. We use it every day.

Capacity

Reusable asset libraries, compute and tool licenses, plus AI partners living in the background — repetitive work and scheduling stop being your problem.

Publishing

Self-developed and self-published; Chess of the Hidden Gates ships in 10 languages, at home and abroad. When the work is done, this system takes it the rest of the way to players.

This is how we grow capacity: one more cell, one more product line.