VibeGame: Prompt-to-Game Development with AI-Native Engine and Self-Evolving Adversarial Agent Team

VibeGame couples an AI-native 2D game engine with an eight-agent adversarial team and training-free experience distillation to turn natural-language prompts into playable game projects.

Building a playable game requires coordinating code, scenes, assets, animations, physics, and configurations, while validating integrated behavior through interactive testing. World models can generate interactive experiences but produce no editable source code, and LLM integrations into commercial engines leave project state scattered across graphical interfaces and binary assets that agents cannot consistently inspect or modify. Many game issues emerge only during play, and fixing one project artifact can introduce inconsistencies across interdependent artifacts. Existing agent systems often reserve runtime checks for final evaluation or limit them to verifying that a game builds and runs.

VibeGame builds on Phaser and represents each game as scenes whose configurations are stored in type-checked JSON, with a scene-node-asset hierarchy and schema validation that detects invalid references before execution. The engine provides frame-synchronous runtime control so agents can pause and resume, advance frames, inspect or modify runtime properties, inject semantic actions, and capture screenshots; the engine's uncompiled source code is copied into the project directory. An eight-role Adversarial Agent Team separates implementation from review and organizes development into intention alignment, parallel development, and adversarial correction, coordinated by an orchestrator. Code sub-tasks pass through architect, programmer, auditor, and player roles, and are complete only after static auditing and runtime play-testing. After acceptance, a training-free self-evolution mechanism distills projects into reusable skeletons, modules, and collaboration contracts for future runs. Evaluation covers prompt-assets-to-game with a provided asset library and prompt-to-game with self-generated art on a GameCraft-Bench subset.

• On the GameCraft-Bench subset with GPT-6.1-Sol, VibeGame in P2G achieves the highest overall score among the listed harnesses, above Codex in P2G. §3.1
• With GPT-5.6-Luna, VibeGame in P2G scores higher overall than Qwen Code in P2G and DSH in PA2G. §3.1
• In the ablation study, removing the AI-native engine lowers the overall score from 53.45 to 43.12, and removing the Adversarial Agent Team lowers it to 48.82; adding evolved experience raises it to 56.37. §4.2
• Under the GPT-5.5 judge, VibeGame with GPT-6.1-Sol reaches 78.57 overall, above Codex in P2G at 77.44, and VibeGame with GPT-5.6-Luna reaches 55.65 overall. §6
• Equipping the native Codex harness with VibeGame's AI-native engine raises its overall score from 64.44 to 67.13, with gains in all four rubric categories. §12
• VibeGame with GPT-6.1-Sol improves over the strongest baseline, Codex in P2G, by 4.19 points. §5

• The paper states that it focuses on 2D games and has not yet validated the framework in more complex 3D scenes or used it to develop complete commercial games that sustain long-term play.
• The paper states that its AI-native design principles are implemented on a web-based 2D engine; bringing them to established engines such as Godot and allowing agents to operate existing game projects remains future work.
• The paper reports that when VibeGame continuously develops games of different types, experience accumulated by self-evolution can instead lower agent performance.
• The paper reports that switching the judge changes both absolute scores and ranking, and that GPT-5.5 has severe judgment flaws as a judge model, so evaluation depends on judge choice.

• The paper defines a game engine as AI-native when it is GUI-independent, runtime-accessible, and source-available, removing barriers for coding agents. §3.1
• Each game is represented as scenes with type-checked JSON configurations and a node hierarchy; shared nodes are standalone definition files, and schema validation can detect missing assets before execution. §3.1
• The frame-synchronous interface lets agents pause and resume execution, advance a specified number of frames, inspect or modify runtime properties such as positions and velocities, inject semantic actions, and capture screenshots. §3.1
• The Adversarial Agent Team has eight roles: orchestrator, designer, artist, architect, programmer, auditor, player, and reviewer; the orchestrator maintains development state across intention alignment, parallel development, and adversarial correction. §3.2
• A code sub-task is complete only after it passes both static auditing with the engine's consistency-checking tools and runtime play-testing with a vision model. §3.2.2
• An independent reviewer acts as the final quality gate on functionality, visual quality, and playability; if the project fails any criterion, the reviewer rejects it with reasons that the orchestrator converts into repair tasks. §3.2.3
• Self-evolution distills accepted projects into skeletons, modules, and contracts; skeletons include error notes and an art pack, modules may include self-checking logic, and contracts adapt the general workflow to recurring features. §3.3
• The test set has 30 tasks from GameCraft-Bench, two from each of 15 game types, and the ablation set is a 10-task subset. §7