Sequence RPG
A 3D RPG where every NPC is a distinct AI agent — spatial awareness, persistent memory, and autonomous in-world actions. The agent layer is feature-complete; the world itself is now being built around it.
NPCs as persistent LLM agents.
Most game AI is a state machine dressed up as personality. Sequence RPG flips it: every NPC is a distinct LLM agent with its own goals, memory, and spatial awareness. They perceive nearby entities, overhear conversations within a radius, and reason about who said what to whom.
The agent layer is feature-complete — one SystemOrchestrator drives 20+ NPC archetypes via a composable prompt-component system. The current focus has shifted to the world itself: physics, villages, terrain, shaders, and the map they all live in.
Four systems that make it feel alive.
One orchestrator, many personas
A single SystemOrchestrator drives every NPC via swappable prompt components. Personas hot-swap without rebuilding the agent loop.
NPCs see, hear, and reason
Perception system feeds nearby entities, overheard dialogue, and location/time context into every decision — so characters act in-world, not in a vacuum.
They remember everything
Per-NPC daily session files with persona-filtered history. Talk to a shopkeeper today; he greets you by name tomorrow and recalls the haggle.
Structured responses that don't break
4-method JSON extraction pipeline with automatic repair for malformed LLM output — reliable structured actions, even when the model improvises.
Building the world.
No concept art yet — I'm building assets directly. Early worldbuilding renders: terrain, character studies, and atmospheric tests.
Where it stands today.
Project kickoff
SystemOrchestrator architected. First working agent loop drives a single NPC through dialogue and simple in-world actions.
Composable persona system
Prompt-component layer lets one orchestrator drive 20+ NPC archetypes. Swapping personas no longer requires touching orchestration code.
Perception + persistent memory
NPCs perceive nearby entities and overhear conversations within a radius. Per-NPC session files persist across playthroughs, filtered by persona.
Agent layer feature-complete
The full AI layer — personas, perception, memory, multi-agent awareness — is working end-to-end. NPCs behave like NPCs.
World & asset building
Designing the world the agents live in: physics, villages, terrain, the map, and the shaders that make it feel like somewhere.
Playable demo
First town wired up end-to-end, cost-efficient inference budget per session, public playtest.