Auditing AI Autonomy in Minecraft: Skill Libraries, Human Intervention, and the Limits of Agentic Progress

We investigate the capabilities and limitations of autonomous coding agents in Minecraft through an independent partial reproduction of a Google DeepMind experiment (arXiv:2610.22319). Four Gemini 3.8 Flash High agents were deployed in survival Minecraft using Mineflayer, a tested library of 54 skills, and a constrained computational budget. The experiment lasted 12 calendar days, including 20.1 hours of active gameplay. Agents executed 3,770 programs, experienced 332 deaths, and encountered a 21.2% program failure rate. A central contribution is a server-log-based auditing methodology that distinguishes autonomous achievements from operator-assisted progress. Sixteen milestones were classified into four evidence levels (A–D), with an additional marker identifying operator-provided helper functions. Agents independently completed the initial technological progression, including diamond tools and obsidian, within 2.4 active hours. Later milestones involved increasing levels of assistance. The final defeat of the Ender Dragon was achieved through operator commands and is explicitly excluded from autonomous agent achievements. The study also documents infrastructure failures, the effects of skill libraries, token consumption, and the challenges of maintaining long-running agent workflows. Our findings highlight the importance of independently verifiable outcomes, transparent reporting of human intervention, and reliable agent-environment interfaces when evaluating autonomous AI systems. This work is an independent investigation and does not claim direct experimental comparability with Google DeepMind's original results.