VibeGame turns a prompt into a 2D game with eight agents

Researchers at Nanjing University, Nanyang Technological University and Liblib.ai have built VibeGame, a system that turns a text prompt into a complete, playable 2D game, and described it in an arXiv paper posted on October 22. VibeGame scored 72.99 on a game-building benchmark, against 68.80 for the best of five other coding agents running the same model. That is a gain of 4.19 points.

The paper had 22 upvotes on Hugging Face daily papers when checked on November 1. It matters because games are a hard test for AI coding agents: a game can build without errors and still be broken when someone plays it.

The problem: a game that compiles can still be broken

Making a game means writing code, laying out scenes, drawing art, setting up physics and animation, and tuning rules so they feel fair. Coding agents such as Claude Code and Codex can write the code. The authors argue that they usually don't test the game by playing it. Many bugs only show up when player input, physics and animation interact in real time.

The team calls the task prompt-to-game (P2G): a text description, and optionally reference images, go in, and a full game project comes out. A second setting, PA2G, gives the agent a library of ready-made art. In P2G the agent has to make its own.

What VibeGame does

The system has three parts.

An engine built for agents. It is based on Phaser, a text-based web game framework. Every scene, character and setting is stored as a validated JSON file, so a missing asset is caught before the game runs. The agent can also pause the running game, advance it a set number of frames, read and change values such as a character's position, press virtual buttons and take screenshots. The engine's source code is copied into each project so the agent can look up how things work.

A team of eight agents that check each other. A designer writes a design document from the prompt. An artist generates the art with image models (the paper names GPT-Image-2 and Nano Banana Pro). An architect plans, a programmer builds, an auditor reviews the files, a player runs the game and judges the screen with a vision model, and a reviewer makes the final call and sends the project back for repairs if it fails. An orchestrator coordinates them. The point of the design is that no agent grades its own work.

Reusable experience. When a game is accepted, the system saves templates, reusable behavior scripts and written procedures for later projects. No model training is involved.

The paper gives an example of the review loop. In one run, the auditor found that a timing indicator existed only as text and had no on-screen band. Another agent found game labels cut off with ellipses, and the reviewer found that two characters stood on the same spot and looked like one.

The results

The team used 30 tasks, two from each of 15 game types in GameCraft-Bench, a benchmark that scores games on core mechanics, content depth, functional visuals and art. Every agent got 24 hours per task. The authors replaced the benchmark's usual judge model, GPT-5.5, with GPT-6.1-Sol after finding cases where GPT-5.5 misjudged games.

The other harnesses were OpenGame, Qwen Code, Kimi Code, DeepSeek Harness and Codex. A blind study with 30 people playing 150 pairs of games found that players preferred VibeGame in 72.7% of judgments on average.

Removing parts hurt the score. On a 10-task subset with the weaker model, dropping the engine cut the overall score from 53.45 to 43.12, and dropping the three review roles cut it to 48.82. The engine alone, added to Codex with GPT-6.1-Sol, raised its score from 64.44 to 67.13.

The gain is larger for the weaker model. With the stronger one, the authors say frontier models already handle core mechanics, and VibeGame's lead came from content depth (4.76 points) and art (7.72 points).

What the paper doesn't show

• Same-model comparisons only. Everything uses OpenAI models and the authors' own runs of the benchmark. They moved it from the Godot engine to Phaser, so the scores are not comparable with the original GameCraft-Bench.
• A small test set. There are 30 tasks, and the ablation uses 10.
• 2D only. The authors say they have not tried 3D scenes or a full commercial game meant for long-term play.
• Self-evolution can backfire. The paper reports that experience built up across different game types can lower performance. The main results were run without it. Experience from the stronger model raised the weaker model's ablation score by 2.92 points.
• Judges disagree. Switching from GPT-5.5 to GPT-6.1-Sol changed scores and the ranking, though VibeGame came first under both. Under the GPT-5.5 judge its lead was 1.13 points over Codex with GPT-6.1-Sol.

The paper's demos include a Sekiro-style boss fight, a Fruit Ninja clone and a Werewolf game with 11 AI players. The paper reports them without scores, so they show range rather than quality.

Where to find it

The paper is on arXiv, with code on GitHub and a project page. A related result on how agents use tools is in our piece on MCP versus CLI.