A Mechanistic Interpretability Analysis of a GridWorld Agent-Simulator (Part 1 of N) — LessWrong