An AI Odyssey, Part 4: Astounding Coding Agents

This article is a personal account from a developer using AI coding agents like OpenAI Codex. The author reports that these agents have become much smarter over the past year, handling a broader range of tasks and providing deeper codebase understanding. They estimate AI assistance rose from 20% to 60% of coding effort. However, the author also discusses limitations such as myopic focus, over-optimization to test harnesses, and potential code debt. The piece explores the balance between leveraging AI for productivity and maintaining human-readable, conceptually consistent code. It also addresses concerns about losing coding skills and the cognitive load of managing multiple agents, offering a nuanced perspective on AI's role in software development.