Build a world.
Let intelligence
live in it.
A player sees a world. A game engine knows what is in it. An agent should be able to work with those same objects and rules.
This is a small, playable experiment in what that could look like: one world, two ways to understand it, and rules for every change.
Walk, jump, explore, and decide what the world needs.
A bridge is a named object with a location and a rule.
Permission, version, validation, then a recorded result.
The Crossing.
There's a ravine between you and the signal flag. Walk to its edge. You can't jump the gap. Then give the world a bridge—and see exactly how that change happens.
STATE / CANONICAL OBJECTS ▾version 0
This is a working browser simulation, not the production MotionVector game engine. Movement, version checks, permission checks, object creation, reversals and receipts are implemented locally in this page. The “agent” is a scripted proposer—no AI model or server is connected. Receipts here are inspectable event records, not cryptographic proof.
LIVE JS / NO BACKENDThere are two ways to meet the bridge. The player can walk across it. An agent can refer to bridge_01 by name, inspect its state, and propose a change. Neither needs a second copy of the world.
The important moment isn't when the bridge appears. It's when the system can answer: what changed, which version it changed from, who was allowed to do it, and what can be undone?
Game engines already know what's real.
A game knows where its player stands, whether a door is locked, which objects can move, and what happens when something collides.
Visual agents may infer some of this from screenshots. Game scripts and APIs can expose it directly. Our question is what happens when those meanings become a consistent interface for humans and agents, with one authority boundary.
A landscape.
Mountains, a ravine, a bridge, a character. A visual representation meant to be seen and felt.
A meaningful world.
Named objects, valid actions, current state, permissions and a history of changes.
An agent can propose a bridge. It cannot give itself permission to build one.
Try revoking world.edit before committing. Try changing the world after a proposal is prepared. The same runtime rejects an unauthorized or outdated action before altering its state. That boundary matters more when an agent is allowed to act without someone watching every frame.
One world. More ways to create and play.
The game is a compact example. The same design questions apply to game creation, intelligent characters, automated testing, multiplayer collaboration and simulated environments.
The proposals below are product directions, not finished MotionVector features.
Describe a change.
Keep the level intact.
A designer asks for a bridge, a new route or a different enemy placement. The agent proposes edits to named objects. The engine checks them against the level rules before they become the next version.
Characters that understand their world.
An NPC can ask which door is open and which path is allowed, then act through a bounded set of verbs. The game retains the authority to reject impossible actions.
Reproduce the exact problem.
A test agent records which version it saw, which operation it attempted and what the runtime accepted. A developer can investigate the resulting state instead of relying on a narrated success report.
Here's what exists. Here's what's next.
There is a difference between demonstrating a principle in a small world and delivering a general game runtime. We want the page to make that difference obvious.
FOUNDATION
The August 2026 architecture snapshot documents DocIR and SceneIR, native rendering, validation, addressed patches, and CLI / SDK / MCP interfaces. These are the established engineering roots of the approach; current release status requires a fresh repository check.
EXPERIMENTS
September context notes describe tactics-game rules and a platformer reference path, alongside agent-facing structured state experiments. They also flag incomplete integration into a single canonical runtime.
LIVE
The Crossing implements a small browser world with movement, named objects, state versions, explicit grants, conflict handling and a local receipt history. It is independently coded and not evidence that the production game pack already provides these behaviors.
DIRECTION
Domain runtimes where player input, agent proposals, tools and simulations can refer to the same canonical state, with permissioned effects and carefully scoped integration into existing engines and operating systems.
Evidence sources: MotionVector architecture snapshot (14 Aug 2026), mvec-os context bundle (23–24 Sep 2026). No live engine repository was available for this page's verification.
Build worlds that can be played, understood, and changed.
A person's imagination defines what the world could be. Its rules determine what can actually happen. We want to give intelligent agents a direct way to participate—without taking those rules away from the people who made them.