Agent reliability
Long-horizon task completion, recovery after tool failure and stable behavior in worlds that keep changing.
Technology
A persistent game agent, running in the open.
VoxelMind is not a chat layer beside Minecraft. The agent observes a live world, chooses one mechanically executable action, receives the real result and carries what matters into the next decision and the next session.
The production loop
Perceive. Decide. Act. Verify. Remember.
One decision-maker, a small executable toolset and an explicit result loop keep behavior inspectable in a world that does not wait for the model.
01
Perception
Nearby blocks, entities, health, inventory, chat and the previous action result become the agent's grounded view of the moment.
02
LLM decision
One model call sees sensory context, recent history, memory and current intent, then chooses exactly one next action and optional speech.
03
Game tool
The chosen action runs through Minecraft mechanics: navigate, gather, craft, place, fight, use containers or wait.
04
Result verification
Success, failure and changed world state return to the loop before the agent decides what to do next.
05
Persistent memory
Shared events, places, relationships and reflections survive the session so a returning player meets the same companion, not a reset prompt.
<100 ms
target reaction time
Fast layer
Immediate danger does not wait for an LLM.
A deterministic reflex layer handles narrow survival reactions such as imminent threats in under 100 ms. It does not plan or replace the agent; it keeps the body alive while the decision loop handles meaning and intent.
Production proof
Measured with real players, sessions and failures.
The same system runs in private hosted worlds and through the Fabric mod in player-owned worlds. Production telemetry covers sessions, decisions, tool outcomes, disconnects and failure modes — the evidence used to improve reliability instead of grading a closed demo.
1,900+
registrations
32/day
new registrations/day · 7-day average
4,800+
CurseForge downloads
2
live surfaces · hosted worlds + Fabric mod
Product metrics through Aug 25, 2026 · CurseForge checked Aug 26
Collaboration
Where the system can be pushed next.
VoxelMind is open to focused technical collaboration where a live game-agent environment is useful.
Evaluation
Production-grounded traces and repeatable scenarios for measuring decision quality beyond a one-off benchmark score.
Voice & presence
Low-latency conversation that stays grounded in what the character can actually see and do.
Scaling
More concurrent persistent agents without losing identity, memory quality or operational visibility.
Building agents for dynamic virtual worlds?
Talk to Nyvoro about reliability, evaluation, voice or production scaling — or meet the current agent in Minecraft first.