An autonomous Minecraft agent designed to reason, survive, learn, and eventually complete the game.
StoneSeed is an experimental game-playing AI that observes Minecraft, selects goals, maintains multi-step intentions, uses executable skills, verifies real world-state changes, remembers outcomes, and responds to emergencies without relying on constant player commands.
StoneSeed explores whether an agent can keep its own intentions alive across a long session instead of only reacting to one command at a time. That makes it a research project about judgment, persistence, and recovery rather than a simple automation script.
The target audience is anyone interested in game-playing AI, agent planning, and systems that can prove they actually changed the world state. It is especially useful as a testbed for the gap between scripted actions and real survival behaviour.
The long-term goal is an agent that can join a fresh server, choose a sensible objective, maintain and resume plans, verify actions, remember outcomes, and steadily complete advancement goals.
Most game bots execute narrow commands or fixed scripts. StoneSeed explores whether an agent can maintain intentions, understand prerequisites, verify its actions, respond to danger, and improve from experience across long gameplay sessions.
The project began as a test of whether an agent could behave more like a planner than a macro.
A runnable agent loop exists as the basis for survival and planning experiments.
The current design emphasizes skills, adapters, verification, and persistence instead of single-command execution.
Opening progression and failed-action recovery are being made reliable enough to benchmark honestly.
Consistent early-game survival before attempting broader advancement planning.
Public language/runtime details are not enumerated on the site page yet.
Agent orchestration and adapter-driven tooling
Observation, decision, and verification loop Persistent intention management Skill adapter layer and reward tracking
Minecraft session runtime with external agent control
Memory and outcome records Project notes and verified state changes
Survival progression checks Manual scenario validation
Minecraft world state Agent skills Project update notes
Minecraft Research environments
The benchmark remains early survival and recovery, not flashy endgame claims.
StoneSeed is still being judged on the opening hours of a session: wood, planks, crafting, shelter, food, and the ability to recover after a mistake. That keeps the research honest and prevents the project from skipping the hardest part of the problem.
View Full Update
A research platform for training strategic agents through imitation learning, reinforcement learning, self-play, and cross-game transfer.
BearPack OpenFront AI uses OpenFront as the first reference environment for a larger general game-intelligence platform. The goal is to build reproducible strategic agents that can learn from demonstrations, compete through self-play, model opponents, plan ahead, and transfer knowledge to...
A fast territorial browser strategy game where every border becomes a tactical decision.
BearFront.io is a real-time world-domination game built around territorial expansion, economy management, contested borders, cities, ports, naval warfare, and short strategic matches.
A deterministic native simulation runtime built for repeatability, replay, rollback, and inspectable behaviour.
Unified Physics Engine is a native physics and simulation project centred on deterministic stepping, state hashing, snapshots, replay workflows, and a stable integration boundary.
Reliable fresh-spawn progression, first-night survival, shelter construction, food acquisition, tool progression, and recovery from failed actions.
