Code2Games offers agent-based world-building, validated in Unreal Engine 5.

Code2Games presents an agent system that generates game worlds from descriptions and verifies that scenes, objects, rules, and scripts work together. The research was published on arXiv on October 4, 2026, starting with a base world in Blender before adapting it for use with Unreal Engine 5.

The problem the team is trying to solve is that each component may appear correct in isolation, but when combined, the game becomes unplayable—for instance, objects in the scene do not match what the script references, or game objectives are not connected to areas the player can access.

Provide test results to fix the game world.

The system uses a shared data representation for scenes and gameplay, preserving relationships between objects, goals, and rules during creation. Once moved into Unreal Engine 5, compilation errors, runtime results, and playtest feedback help resolve inconsistencies.

The team also proposed the GameCode4D assessment kit, which includes 10 fixed game problems evaluating image quality, interaction accuracy, multi-format artifact quality, and playable game quality. The researchers reported superior results compared to the baseline method used for comparison, with these findings falling within the experimental scope of the research.

Code2Games remains a proposal for a system and experimental results, not a ready-made game creation service that general users can immediately activate. The focus of this work is on verifying that various components become a game that actually functions together.

Data: Code2Games on arXiv