A framework helps agents build and test complete 2D games
VibeGame: Prompt-to-Game Development with AI-Native Engine and Self-Evolving Adversarial Agent Team · arXiv:2610.33106
In plain words
VibeGame is a framework that turns a written request, sometimes with reference pictures, into a complete 2D game. Its software agents can inspect the game's files and control the running game to check how it behaves.
Separate agents build parts, test them, and review the finished project, sending rejected work back for repair.
On an adapted game-creation test, VibeGame scored higher than Codex when both used the same game-generating model. The result applies to that test, not to every kind of game.
The research
A game can build successfully yet still fail when its code, visuals, and controls interact during play. It remains unclear whether an agent can inspect those interactions and revise a connected project, rather than stopping when it has produced files that appear structurally sound.
VibeGame combines an agent-friendly game engine, separate building and review roles, and reusable project experience to create and check complete 2D games.
In a modified GameCraft-Bench test of prompted game creation, GPT-6.1-Sol judged demo recordings against each task's requirements; researchers averaged three passes.
With the same generating model, VibeGame scored 72.99 versus Codex's 68.80 on a weighted 0–100 rubric, which assigns zero if a game cannot be built or played.
The tasks, time budget, and judging protocol were held constant across those two agent harnesses. The result supports an advantage for VibeGame's development setup on this adapted test.
The evaluation covers 2D games on tasks adapted to Phaser, so it does not establish how VibeGame performs on 3D games or complete commercial games designed for long-term play. The component and experience comparisons also apply only to their reported ablation conditions.
What the evidence shows
- A game must work together
VibeGame is a framework that turns natural-language requests, optionally with reference images, into complete 2D games. The problem is that a game depends on connected parts such as code, scenes, art, animation, and physics, and some defects appear only during play.
Checking that files are present or that a game builds does not by itself show that these parts work together.
- Expose files and runtime
VibeGame's engine makes project parts and their links available as structured text, with checks for invalid references such as missing assets. Agents can also pause or advance the running game, inspect properties, inject actions, and capture screenshots.
This gives them a way to examine behavior during play, where interactions among input, physics, animation, and game logic can reveal problems that file checks miss.
- Build, challenge, revise
The Adversarial Agent Team divides implementation and review among eight specialized agents. An auditor checks project structure, a player agent runs the game, and a separate reviewer assesses the complete project.
When the reviewer rejects it, the orchestrator assigns repairs and the team repeats the cycle. Accepted projects can also supply reusable templates, behaviors, and procedures for future runs, which are adapted and checked again.
- Compare the game projects
GameCraft-Bench tests game creation from task descriptions. The researchers adapted 30 tasks across 15 game types to Phaser.
On prompt-only tasks, GPT-6.1-Sol judged recorded demos against task-specific requirements; the weighted overall score combines rubric categories and becomes zero if a game cannot be built or played. VibeGame scored 72.99 versus Codex's 68.80 when both used GPT-6.1-Sol as the generating model.
- Probe parts and experience
The advantage appeared across most tested game types: VibeGame exceeded the strongest corresponding-model baseline in 13 of 15 types, though this does not mean it won every task in those types. Separate comparisons on a 10-task subset found lower scores when the engine or three review roles were removed.
Adding experience built by GPT-6.1-Sol on five other tasks also raised GPT-5.6-Luna's score on that subset, supporting transfer in this specific setup.
Deep
- Original paper authors
- Wenbo Hu, Ken Li, Jiazhe Wei, Yukang Cao, Weiyi Hong
- Submitted to arXiv
- License
- CC BY 4.0