top of page
Crafting immersive experiences that bridges players and game worlds.
Gameplay and Engine Dev
AI-Agent Game Development Workflow
This project explored how LLM-based agents could be integrated into real-world gameplay engineering workflows inside a production Unity ECS environment. The workflow was designed as an autonomous gameplay development pipeline capable of planning tasks, generating gameplay implementations, constructing simulation-layer unit tests, and validating gameplay behavior without direct human intervention.
By enforcing deterministic validation workflows before implementation began, the system significantly improved implementation reliability for gameplay-related engineering tasks and achieved over 80% autonomous implementation accuracy for non-architectural gameplay features.
Workflow Architecture:
The workflow operates through a staged execution pipeline coordinated by a centralized orchestration layer. Once a gameplay request is received, the system first analyzes the task and retrieves relevant ECS systems, gameplay components, and runtime data structures from the project codebase.
Before implementation begins, the workflow generates simulation-layer unit tests that define the expected gameplay behavior and validation criteria. These tests are executed independently to ensure the validation environment itself is deterministic and reliable. Only after the tests pass validation does the implementation stage begin.
The implementation agent then generates gameplay code, compiles the project, executes all unit tests, analyzes failures, and iterates autonomously until all validation conditions are satisfied.
Technical Challenges:
1. Interpreting Data-Driven Gameplay Architectures
Problem:
Large portions of the gameplay framework relied on externally-authored spreadsheet configurations that defined ECS components, gameplay behaviors, balancing parameters, and runtime gameplay schemas.
Because these gameplay structures existed primarily outside the codebase, the agents were unable to reliably infer gameplay relationships or understand runtime data layouts directly from raw spreadsheet content. This frequently resulted in incomplete gameplay implementations and inconsistent usage of gameplay data.
Solution:
The workflow integrated the Luban data pipeline to automatically generate strongly-typed runtime C# definitions from gameplay schemas. Instead of reasoning about spreadsheets directly, the agents analyzed generated runtime code structures and ECS-facing APIs inside the production codebase.
Additional Python tooling was introduced to preprocess gameplay data, inspect generated schemas, and expose structured runtime information during planning and implementation stages.
Result:
This significantly improved gameplay structure comprehension and allowed the workflow to generate gameplay implementations that aligned correctly with existing ECS systems, runtime data layouts, and production gameplay architecture patterns.
2. Autonomous Gameplay Validation
Problem:
Gameplay systems depended on runtime ECS initialization and world ticking behavior that traditional unit tests could not reproduce reliably
Solution:
Generated simulation-layer unit tests capable of:
• ECS world reconstruction
• Gameplay system registration
• Component attachment
• Deterministic world ticking
Result:
Enabled autonomous compile → test → validate → iterate workflows entirely inside the gameplay simulation layer.
bottom of page