Kaggriculture is a two-player Kaggle competition: farm a shared board, sell into a moving market, out-earn an opponent over 720 turns. I built a simulator that matches the real engine exactly, then a planner that searches ahead through it instead of reacting turn by turn.
A reactive agent can't see the trap it's walking into.
Land, water, harvest, sell, repeat: an agent that only reacts to the current turn buys land too late, plants the wrong crop for the days left in the season, or gets outpaced by a competitor at the market.
Simulate first, then search through the simulation.
A deterministic simulator reproduces the game's rules, checked turn by turn against Kaggle's own engine. A planner searches move sequences through it, models the opponent, and hands off to an emergency handler near the end of the season.
Tests assert the simulator against Kaggle's real engine turn by turn, so a planner search through it searches the actual rules, not an approximation.
The planner earned its search budget one phase at a time. Ten phases in three days, each one validated before the next began.
Phase 2 checked the price formula against the live engine before Phase 3 trusted the simulator to search through. Phase 5's planner had to beat every earlier agent before it became the baseline, and Phase 9 replaced a discount-formula guess at the opponent with rollouts of its own policy.