Multi-Unit Coordination Dominates Crop Choice: An Empirical Study of a Heuristic Agent in the Kaggriculture Resource-Allocation Environment

We study heuristic agent design for Kaggriculture, a two-player, 720-turn farm-management simulation combining production scheduling, labor allocation, land expansion, and an inventory-driven dynamic pricing market. The question we ask is practical: which design decisions actually move the needle, measured against controlled baselines instead of a single demonstration run. We build a baseline ladder of four agents of increasing sophistication: a random policy, two single-tile crop loops (carrot and wheat), and a multi-unit "wheat swarm" that coordinates a farmer and several hired hands across many tiles with land expansion and a market-aware selling rule. We evaluate each across ten fixed random seeds against a common opponent. Multi-unit coordination and land expansion produce a large, statistically overwhelming effect (mean final money 6,304 USD vs. 3,508 to 3,709 USD for single-tile agents, Welch's t up to 11.8, p < 0.001, Cohen's d > 3.5); crop choice alone (wheat vs. carrot, both at single-tile scale) is a small but, on formal testing of the raw per-seed data, statistically real difference that is an order of magnitude smaller in effect size than the coordination effect, and confirmed nonparametrically with a Mann-Whitney U test. A paired ablation of a simple market-timing selling heuristic shows a small positive mean effect (+126 USD) with a standard deviation an order of magnitude larger (1,195 USD, sign flips in 5 of 10 seeds; paired t = 0.33, p = 0.75, Wilcoxon p = 1.00). This result is formally inconclusive and consistent with the environment's own damped glut-side pricing function for wheat. We describe the data pipeline used to turn episode runs into the per-seed results these comparisons are built on, document two implementation pitfalls specific to this environment (a file-loader ambiguity and a unit-coordination bug that silently nullified hired labor), and scope fertilizer use, livestock income, opponent modeling, and planning or search methods as untested future work.