Let a language model drive game agents — NPCs, companions, factions, colonies — constrained to actions you explicitly register. The model picks a behaviour; validated, deterministic game code carries it out.
- In a shipped game — inside the game itself, on the player's machine, so players install nothing. The plan is to build on LLMUnity.
- During development — Ollama, to swap and benchmark models in seconds. Players never need it.
- Optionally — cloud APIs, behind the same provider interface.
Wiring a model into a game is not the hard part. Making its output safe to execute is.
In prototype measurements, a 2B local model chose the correct game action 58.5% of the time when the rules lived in the prompt. With a state-derived schema, constrained sampling and a deterministic validation layer, the same model reached 95%, and beat a 4B model that had none of that scaffolding.
The scaffolding is the framework. The model is a swappable part.
- Illegal actions and unknown targets are removed from the sampling grammar, so the model cannot emit them.
- The executor re-validates every choice independently, so a provider without constrained decoding is still contained.
- The model never touches Unity. It chooses; your code acts.
Early development, not yet ready to drop into a game. The framework is being built phase by phase from that measured prototype.
- agenerela/agenerela — the framework
- Build plan — architecture, phases, and the evidence behind each design decision
- Design — what a developer and a player will see, screen by screen
- Contributing — how to open a pull request
Senior design project, California State University, Northridge — COMP 490 / 491.
