A SonarQube cyclomatic complexity flag on a 30-line file shouldn't be a crisis. But when your AI coding agent wrote the file — cramming entire components onto single lines, over-engineering edge cases you never specified — a routine CI failure becomes an epistemological question: do you actually own this system? The author, a practitioner with a dialed-in agentic workflow using OpenAI's Codex, discovered that their agent had produced code that was correct, well-covered, and completely opaque. The complexity was warranted by business requirements. The agent happily iterated on it. But no human had ever reviewed the logic before the linter complained. The fix was trivial — tell Codex to refactor for SonarQube — but the precedent is not. This is code that existed in production without a human mental model of its internals. The essay's central metaphor is divebots: autonomous mining robots that build underwater operations while you sit on the boat. The Optimists say collect your gold nuggets. The Skeptics warn about grotesque undersea cathedrals nobody can maintain. The author rejects both camps and lands on a third position: you need fundamentals deep enough to detect quiet failures, breadth enough to build an abstraction map of any system, and the ability to dive to arbitrary depth just-in-time when something breaks. This is a direct challenge to two popular prescriptions circulating in the industry. The "read every line" camp demands a level of oversight that is economically impossible at agent output speeds. The "just vibe-code and let the models handle it" camp assumes failure modes that don't exist yet won't emerge — a bet against every complex system in history. Neither scales. Neither is honest about the economics. The most uncomfortable paragraph in the piece is the one about junior engineers. The traditional apprenticeship — years of hand-coding, debugging, production incidents — requires employer patience and economic subsidy that the market will not provide when agents can ship features in hours. The author is blunt: "If you think that the industry at large will adopt careful code reviews, I've got a bridge to sell you." The training pipeline that produced today's senior engineers is not available to the next cohort. What replaces it? The essay argues for three pillars: fundamentals (understanding failure modes, networking, concurrency, data structures), breadth (system-level abstraction maps that let you navigate without having written every module), and just-in-time depth (the ability to cold-dive into unfamiliar code and fix it under pressure). This is the skillset of a good SRE or CTO, not a line-level coder. The argument is that agentic tooling makes this the baseline, not the ceiling. The piece cuts off before fully developing the "fundamentals" argument, which is its weakest point. But the diagnostic is sharp: the industry is building systems at unprecedented speed with unprecedented opacity, the abstraction layers are getting thicker, and the question of who actually understands what's running in production is no longer academic. It's a staffing, liability, and resilience problem that nobody has solved.
