747s and Coding Agents

In a widely shared blog post titled '747s and Coding Agents,' AI lab engineer Carl Kolon draws a striking parallel between modern software developers and veteran 747 pilots. He recounts a conversation with a pilot who lamented that after decades of flying, 'there's no improvement'—a fate Kolon now sees for developers as AI coding agents (like Claude Code, GPT-4's coding features, or GitHub Copilot Agents) handle increasingly complex tasks. Where developers once had to deeply understand systems through debugging and research, they now often delegate entire features to AI, from adding pagination to fixing complex bugs. This shift means engineers skip the learning process that traditionally came with problem-solving, potentially stunting long-term skill growth.

Kolon details how the workflow has evolved: from using LLMs as enhanced search engines to today's agents that implement features 'one-shot' with minimal human intervention. While this boosts short-term output, he warns it creates a '747 problem'—developers may work for twenty years without becoming more skillful, as they don't internalize the knowledge needed for good design decisions. The core issue isn't prompting (which is getting easier) but the erosion of hard programming knowledge. As agents handle bigger tasks, developers who do step in face unfamiliar, AI-generated code that's 'a little bit wrong,' leaving them lost. The post has sparked debate about whether AI is augmenting engineers or creating a generation of technical managers who can't build without AI assistance.