Status: In progress.

Objective

Build an agent for a two-player farming sim that ends a 30-day season with the most money.

Approach

A modular agent, split so each piece can get smarter without breaking the others:

Module	Job
State parser	Turns the raw observation into something the agent can reason over
Priority task executor	Picks the most valuable action available this turn
Market order book	Tracks prices and decides when to buy and sell

Strong heuristics plus bandit-tuned parameters before any RL. A 30-day season is short, and a well-tuned rule-based farmer is a baseline I'd want to beat before training anything.

This is also a small version of the decision problem from Gaby and Tom: an agent balancing what it already knows works (exploit) against trying something new (explore).

Where it's at

• ✅ Module skeleton
• 🔄 Self note: the current agent is a stub that plants wheat on every tile. It's a placeholder, not a strategy.
• ❌ Bandit tuning of heuristic parameters
• ❌ RL