How VibeGame Rebuilds a Game Engine for Self‑Evolving AI Agents

Background and Challenge

Large‑language models have made one‑line code generation commonplace, but existing demos stop at a runnable version, treating that as the final goal. In real game development the first playable version is only the starting point, and traditional agents face four critical gaps: inability to generate art, inability to keep up with 60 fps, lack of engine accessibility, and difficulty understanding design requirements.

Prompt‑to‑Game Development

The research team from Nanjing University (PRLab) and NTU proposes the task Prompt‑to‑Game Development : from a natural‑language description and an optional concept image, deliver a complete, playable game project that remains editable.

AI‑Native Game Engine

To make the engine fully readable and writable by agents, VibeGame builds on the 2D framework Phaser and redesigns it with three key properties:

All assets are text. Scenes, animations, collisions and parameters are stored as type‑checked JSON files, so agents edit game content the same way they edit code, and low‑level errors (e.g., missing textures) are caught before execution.

Frame‑level pause and step. Agents can pause the game, advance frame by frame, inject actions such as "jump" or "attack", then read back state and compare screenshots, turning a real‑time system into a controllable test harness.

Source code is exposed. The full engine source is shipped with the project, allowing agents to read implementation details when documentation is insufficient.

These designs make the engine the first that is completely controllable for both code and game modification.

Eight‑Agent Development Team

VibeGame assembles eight specialized agents—producer, planner, artist, architect, programmer, auditor, tester, and verifier—forming an adversarial team. The core principle is that any output from one agent must be validated by a different role that did not participate in its creation.

Three‑Step Workflow

Intent alignment. The producer queries the user, the planner writes a design document, and the artist creates a concept image; both are approved by the user and become the shared specification.

Parallel development. Code, art, and design streams progress simultaneously. Code is first planned, then implemented, and must pass two checks: audit against the design doc and in‑game testing to verify behavior and visuals.

Adversarial correction. After the three streams finish, the verifier conducts a final review of functionality, visuals and feel. Any failure is returned with a reason, turned into a repair task, and re‑enters the pipeline until it passes.

The process is fully automated, though users may intervene at any point, feeding feedback back into subsequent iterations.

Self‑Evolving Assets

Each delivered game contributes three reusable assets to a growing library:

Skeleton. A runnable template preserving project structure and tuned parameters for a game genre.

Module. Cross‑project components such as health bars or mini‑maps, each bundled with its own verification logic.

Contract. Documentation of role responsibilities for a given feature (e.g., what the planner, programmer, artist and tester each deliver for a map).

When a new project starts, relevant assets are retrieved as a baseline but must be re‑adapted and re‑validated, ensuring continuous improvement without updating any model parameters.

Conclusion

VibeGame succeeds by (1) rebuilding the engine from the ground up to be fully observable and controllable, (2) structuring eight agents in an adversarial team that self‑test and self‑criticize, and (3) preserving experience as reusable skeletons, modules and contracts, so each subsequent game starts from a higher baseline. This closes the loop of development, testing, iteration and evolution, turning “one‑sentence game generation” from an endpoint into a true starting point for anyone who wants to build games.

Paper: VibeGame: Prompt-to-Game Development with AI‑Native Engine and Self‑Evolving Adversarial Agent Team (https://vibegame.tettet.org/technical_report.pdf). Repository: https://github.com/tettethu/VibeGame.