Skip to main content

Project Roadmap & Goals

Our immediate goal is to demonstrate a stable "NPC Tool-Loop"—a system where a Minecraft NPC can perceive its environment, make decisions using an LLM, and execute those decisions reliably.

The Minimal Viable Proof​

We are building a scenario where two NPCs, NPC_A and NPC_B, must interact to solve a simple problem:

  1. Perception: NPC_A observes the world and sees NPC_B.
  2. Decision: NPC_A decides to approach NPC_B and ask a question.
  3. Social Awareness: If NPC_B is busy (e.g., crafting), NPC_A must decide whether to wait or try again later.
  4. Conclusion: NPC_A records the result in its memory and the session transcript is saved.

Core Tool Set​

To keep the initial research focused, we use a minimal set of validated tools:

  • observe(): Gather data about surroundings.
  • move_to(target): Navigate to a specific actor or coordinate.
  • say(target, text): Communicate with another NPC.
  • wait(duration): Pause for a short period.
  • remember(note): Store a key fact in local memory.

Future Milestones​

Once the basic tool-loop is proven stable, we plan to expand into:

  • Shared Economy: NPCs managing shared chests and crafting dependencies.
  • Role-Based Societies: Clear divisions of labor (e.g., Gatherers vs. Crafters).
  • Survival Challenges: Introducing scarcity and hostile entities to observe emergent cooperation.
  • Dynamic Skill Generation: Allowing NPCs to "learn" new complex behaviors over time.

Success Criteria​

A simulation run is considered successful when:

  • Multiple bots join and stay connected to a headless server.
  • The transcript shows logical turn-taking and tool usage.
  • The agents successfully navigate social state (e.g., waiting for a busy partner).
  • A structured data artifact is produced for every run.