Weekly ranking, Technology
Agents in Games
90Original Thinker
Original multi-agent game experiments with concrete logs showing arbitrary hierarchies, compounding caution and contrasting GPT and Claude swarm styles, though sample sizes are small.
What it adds
Original experiments pitting frontier model swarms in Minecraft and Mindustry, quantifying emergent leadership by directory order, cascading caution, and distinct GPT vs Claude coordination styles, with open-sourced harnesses.
Claim by claim
- 16 the author's own
- 0 an AI would say
- 0 known elsewhere
- 3 common knowledge
Its strongest original claims
- In a 50-agent GPT-6.1 Sol Minecraft civilization run, agents became increasingly risk-averse after a first-night wave of deaths and the colony never thrived.
- 22 of the run's 34 agent deaths occurred in the first night.
- Leadership emerged from directory order: 46 of 49 agents first messaged Ada, the first name, who received 616 DMs with no vote.
- Agents showed a tendency toward bureaucracy, e.g. four agents spent 11 minutes appointing a sole fertilizer operator for one bone meal.
- After a tossed-food incident, mentions of 'toss' rose from 7 messages to 125, 85 of them warnings.
Scored by Expulp Verdict for what the piece adds beyond what is already known and published. The full verdict shows every claim with its reasons. Leaderboard, what Verdict gets right and where it is weak.