tl;dr: A Haskell developer argues against letting LLMs write your code, since it erodes both codebase quality and programming skills, and instead recommends using agents for planning, research, review cycles, and boring cleanup tasks while you remain the primary coder. He also advises treating token limits as service outages to prepare for, avoiding frontier models when possible, and never sending fully-generated PRs or communications to human collaborators. He claims this approach roughly doubles his productivity while preserving the joy of programming.