This Minecraft dragon-slaying agent shows what honest AI game demos look like

A Minecraft agent that beats the Ender Dragon in under nine minutes is interesting. What makes this repository worth examining is the architectural honesty about what the models actually do — and what they don't.

The planner (GPT-6 Astra or GPT-5.6 Sol) sets objectives and waypoints. JEV selects one action from a bounded legal set. Mineflayer handles pathfinding and the game protocol. The models never see raw screenshots or issue individual key presses. They operate on structured game state: block positions, inventory contents, coordinates from a known seed. That distinction matters. A lot of claimed "AI plays game" demos are really "AI watches game and narrates" — this one is wired into the actual action loop through typed decisions.

The latest full run, nether-final-08, completed in 8 minutes 43 seconds, down from 14 minutes 31 seconds in the previous recording — a 40% reduction. The End combat stage alone dropped from 332 seconds to 152 seconds. The agent started with an empty inventory in a fresh Survival world, used the Nether for travel, and killed the dragon with six bed explosions on the first landing, finishing at full health with no deaths. That sequence required 131 JEV decisions and 35 planner calls.

The seed selection is candid: the repository describes it as a useful speedrun seed, not the easiest possible seed. The portal is naturally active, three supply chests contain 21 obsidian, and the route skips blaze rods entirely. Known coordinates are supplied to the models — the agent isn't discovering the world, it's executing a surveyed route with model-guided decision-making at each step.

The evidence chain is unusually rigorous for a personal project. Victory requires a victory.json file backed by dragon-death evidence — either the kill advancement or server-confirmed zero health during the dying phase — plus the exit portal event and a check of the world DragonFight state. The combat sensor is a read-only Java observer that reports dragon head position without touching game rules or entity state. The README explicitly flags that combat-probe.mjs must never be presented as a model-controlled run.

The recording infrastructure is also notable: a hidden native Minecraft client renders the game from a local protocol mirror, capturing native entity models and animations rather than a third-party viewer. The overlay strip shows model names, elapsed time in milliseconds, XYZ coordinates, the current planner objective, and the selected action — all from recorded responses and game events, not live annotation.

The main limitation for anyone wanting to reproduce this is practical: the native renderer and launch scripts were developed on macOS, credentials route through Google Secret Manager with a specific project configuration, and the actual run recordings are excluded from the repository. You get the code, tests, route data, and a detailed evidence methodology — but the world and video stay local.