I built a persistent Minecraft agent system that grew from a single ChatGPT-controlled bot into five local-model agents. My work spans the TypeScript decision loop, shared world context, tool execution, telemetry, and an experimental evaluation track.
Coordination and tools
Each agent specializes in exploration, farming, mining, building, or combat. A shared Team Bulletin puts teammates' actions and positions into each prompt, while a central resource stash lets the agents exchange materials. Coordination uses shared context rather than a dedicated coordinator bot.
An event-driven brain separates strategic, reactive, critic, and chat decisions. Actions pass through role-specific tool gates. Hand-written TypeScript skills coexist with Voyager-style JavaScript skills and generated skills. Mission Control, per-bot viewers, and recorded trajectories make behavior inspectable.
Evidence and engineering boundaries
The public repository documents both normal gameplay and a separate reproducible research track. Preserved world fixtures and server-checked outcomes help distinguish a client's apparent success from what happened in the game. Successful controls and failed qualification attempts are retained.
The advancement ledger is a historical scoreboard: its published 42/122 record is a September 28, 2026 snapshot, not a live progress feed.
What remains unproven
The research track's scripted collection trials and fake-server transport tests qualify infrastructure. They do not demonstrate model learning, cost savings, or robotics transfer. A model pilot remains separate work. Reliable long-horizon coordination is still an engineering problem, not a solved outcome.
Voyager skill library from MineDreamer/Voyager.
