How it works

The agent does the plumbing. People decide at the gates.

How a team builds a game with gejm.dev, gejm.dev Bible and their own agents. Agents draft, check and reproduce. People own the canon, the art, the code decisions and the verdicts.

Who does what

Writer, narrative designer

gejm.dev Bible and its studio chat

Canon: characters, places, props, lore, the timeline.

Art director

gejm.dev Bible

Style records, the world style guide, accepted concept art.

Game designer

gejm.dev Bible, KiwiPow Server

Features and their rules, gameplay numbers.

3D or 2D artist

Blender, the engine's viewport

The look of every model and painted layer.

Developer

KiwiPow IDE or any agent client

Scripts, rulings on the code, the proof.

Tester, producer

Playtests, KiwiPow Server

Verdicts and playtest notes.

Agents

MCP to the bible and to their own engine

Drafts, proposals, checks, reproductions and fixes. Never canon, never a verdict.

The flow

  1. Canon

    Canon first

    Writers and the art director

    The world lives in gejm.dev Bible as records of kinds the team defines: characters, places, factions, props, styles. The agent researches and drafts; every change it offers is a diff a person accepts. A guard's height, a city's palette or a character's voice is decided once.

  2. Plan

    A feature is asked for and planned

    Game designer

    Someone writes a feature in their own words. With KiwiPow Server it becomes an approved spec that cites bible records: scenarios, one provable rule per line, named open questions. Without KiwiPow, the developer and their agent write the spec in the game's repository.

  3. Split

    Claim, map and split

    Developer

    The agent reads the spec, the records it cites and the game's scripts, and raises every place the game disagrees with the spec. The developer rules. The work is split into slices: a feature folder of scripts, a set of models, a room.

  4. Build

    Build in parallel, each session proving its own work

    Agents

    Each slice gets its own agent session and its own headless engine. A half-written script in another session cannot break it. In each one the agent writes TypeScript, sets up the situation in one call, steps the world and drives the player, queries what happened, walks a system line by line when something surprises it, and pins the behaviour with a test.

  5. Draft

    Drafts for artists, checked on every export

    Agents, then artists

    The agent drafts models at gameplay scale with surface roles, colliders and sockets working. The artist finishes the ones worth keeping in Blender; each export is checked against the game and reloaded live, and a broken export keeps the last good version in the world.

  6. Play

    Play it

    Testers

    A person plays the running game. A playtest note carries a replay, so the agent reproduces exactly what the tester saw, fixes it, and proves the fix with a test that replays the note.

  7. Accept

    Accept and close the loop

    Testers, then writers

    The tester judges behaviour by behaviour from the tests that prove it. What the build taught about the world goes back to the bible as a proposal: a door that had to widen, a character's final height. A person accepts it, and the next feature starts from updated canon.

Worked example: ten props for a market square

  1. Brief

    The bible says what each prop is

    The agent lists the props from the place's record and fetches, for each, the record, its art direction, the accepted reference images and the style guide.

  2. Fan out

    Ten sessions start

    One per prop, each with its own headless engine.

  3. Draft

    Each agent drafts at the game's real scale

    A market stall the character can stand at, a cart with a socket for its goods, a well with a collider the player cannot walk through. It bakes, views the asset from fixed angles, and fixes what the checks report until the ready verdict is clean.

  4. Review

    The artist sees ten drafts side by side

    Artist

    Each next to a character-sized reference figure, with its triangle count against the budget and its gameplay checks green.

  5. Hand over

    The keepers get a starting Blender file

    Artist

    Surface roles and gameplay elements are already in it. From then on the artist's file owns them.

  6. Model

    The artist models properly

    Artist

    Each export reloads in the running game and reports, against the object it lies in, anything that broke: a flipped face, a door too narrow, a budget exceeded.

  7. Record

    The bible learns the model exists

    The finished model is proposed back to the prop's record, so the next agent knows it is there.

The artist spent their time on the look. Scale, collision, sockets, budgets and export settings were never their problem.

Worked example: a playtest note

  1. Stuck

    A tester gets stuck between a crate and a wall

    Tester

    They take a note in the viewport, where they stand.

  2. Replay

    The agent replays exactly what the tester saw

    The note carries the session's replay. The developer's agent replays it in its own engine and stops at the tick the tester stuck.

  3. Cause

    It queries the world instead of guessing

    It sees the crate's collider overlapping the walkway, and traces it to the crate's model.

  4. Fix

    It fixes the generator, or tells the artist

    If the artist's file owns the crate, the agent names the object to move. Either way it adds a test that replays the note and asserts the player passes.

  5. Proof

    The note is resolved with the test as its proof

    Tester

What makes it fluent

One source of truth per kind of fact

The world's facts live in the bible, the build's rules in specs, behaviour in tests, the look in the artist's file. Nobody copies a fact between them by hand.

Every handover is text

Records, specs, scripts, generators, test results and reports are all text an agent reads and a person reviews as a diff.

Every agent proves before it reports

A session never says “done” without a test, a judge or a ready verdict behind it.

People decide at the gates

Canon changes, spec rulings, art and verdicts belong to people. The agent proposes at each gate and waits.

Not in the flow yet