Eight-Agent VibeGame Builds a 2D Game From a Prompt and Beats Comparable Harnesses on GameCraft-Bench

Teams from PRLab at Nanjing University and S-Lab at Nanyang Technological University have published VibeGame, a framework that creates a complete 2D game project from a text description and optional reference images, including code, scenes, animation, physics and configuration. Inside, a team of eight specialized agents does the work, and the project itself lives as text files that an agent can read and edit without a graphical editor. On a subset of GameCraft-Bench, in the mode where the system draws its own graphics, it outperformed other agent harnesses built on the same base model, including GPT-6.1-Sol.

Why games are harder than ordinary code

The authors start from the premise that a game is one of the most demanding kinds of software. It must sustain an interactive world in real time, hold a single visual style and follow balanced rules. Many bugs show up only during play: player input, physics, animation and interaction logic combine in ways that break a formally correct project on the fly.

In the authors' assessment, existing approaches cover the task only partially. World models such as Genie generate interactive experiences from video but do not produce editable source code. LLMs inside commercial engines, such as DreamGarden for Unreal, run into project state scattered across graphical interfaces and binary files. And many agent systems check the game only at the end, or merely confirm that it builds and launches.

An engine that agents can read

The foundation is the web framework Phaser, in which projects are expressed entirely as text. To keep agents from creating chaos, the authors added structure. A game is a set of scenes, configured in type-checked JSON files. Each scene is a hierarchy of nodes, and a node combines appearance, collision geometry, behavior scripts and settings. Shared nodes are stored in separate files. The schema catches broken references, for example to a missing asset, before the game even runs.

The second feature is frame-by-frame control. An agent can pause the game, advance it by a set number of frames, read or change the position and velocity of nodes, issue a semantic command such as "press jump" and take a screenshot. Testing therefore runs at the agent's pace rather than in real time: frame by frame, with scripts or bots. The third feature is that the engine's full, unminified source code is placed in the project directory, so the agent consults the exact APIs rather than outdated documentation.

A team that argues with itself

The roles are split as follows: orchestrator, designer, artist, architect, programmer, auditor, player and reviewer. The designer turns a request like "a boss fight in the spirit of Hollow Knight" into a design document covering mechanics, progression and numerical systems. The artist uses image generation models, including GPT-Image-2, to produce concept art and assets, remove backgrounds and normalize sizes. Once the user confirms, the document and concept art become the shared reference for everyone.

Code passes through a chain of architect, programmer, auditor and player. The auditor statically checks the result with the engine's consistency tools. The player launches the game through the runtime interface and uses a vision model to judge whether the picture matches the intended design. A subtask counts as done only after both checks. At the end, an independent reviewer evaluates the project on game behavior, visual completeness and consistency, and on how the controls feel. If something fails, the reviewer sends the project back with reasons, and the orchestrator turns them into rework tasks.

Self-evolution without training

After a project is accepted, the system extracts three kinds of reusable experience from it: skeletons, modules and interaction contracts. A skeleton is a runnable project template for a genre: it has no assets, but includes the code architecture, tuned values and scene flow. New projects adapt and re-verify these materials rather than copying them blindly. Anything unverified or project-specific stays isolated. The model weights do not change in the process.

Ablations show that removing the AI-native engine or the adversarial reviewers significantly lowers the overall score. Experience extracted by a stronger model helps improve the results of a weaker one.

What to keep in mind

Evaluation uses a subset of GameCraft-Bench along four axes: core mechanics, content depth, functional visuals and artistic presentation. The final result is a weighted score from the paper's appendix. The available text gives no specific values, so the size of the advantage over competitors cannot be verified. The authors themselves ran the comparison with other harnesses. The framework is limited to 2D and tied to Phaser. For developers, the takeaway is practical: a text representation of the project, frame-by-frame runtime control and an independent reviewer role are techniques that can be carried over to other agent pipelines. The code and demo are published on the project website and in a GitHub repository.