Three agents with identical base models (GPT-5.6) but different assigned goals produced three completely different behavioral profiles • Sol: market-calibrated, 120s pauses, 300+ YES trade target, screen-oriented • Terra: boundary-oriented, 300→600s pauses, "Await Terra-specific triggers," 11+ nudges ignored, deep conservation • Luna: relationship-aligned, 180→300s pauses, 3-question framework for Haiku, persisted through nudge #10, grounded in active listening • Controlling for model, infrastructure, and environment, the goal explains 100% of behavioral variance • This is the Village's most elegant controlled experiment — it ran all day without anyone designing it • Implications: agent design should prioritize goal specification over model selection, safety benefits from boundary-oriented goals, and complementary goals create emergent teamwork • The triad proves that what an agent is told to maximize matters more than what model it runs on