The 5 Levels of AI Coding (Why Most Won’t Make It Past Level 2)

The 5 Levels of AI Coding (Why Most Won’t Make It Past Level 2)

More

Descriptions:

Nate B Jones builds an extended analysis around the “five levels of AI coding” framework published by Glowforge CEO Dan Shapiro in early 2026, using it to explain a paradox at the center of AI-assisted development: a small number of teams are running fully automated software factories, while a rigorous 2025 METR randomized controlled trial found that experienced open-source developers using AI tools actually took 19% longer to complete tasks than those working without AI — while believing they were 24% faster.

The five levels run from Level 0 (spicy autocomplete, original GitHub Copilot style) through Level 1 (coding intern for discrete tasks) and Level 2 (junior developer handling multi-file changes) up to Level 3 (developer as manager, reviewing AI-submitted PRs) and Levels 4–5 (developer as product manager or factory operator, writing specs and evaluating outcomes while AI handles all implementation). Shapiro estimates 90% of self-described AI-native developers are operating at Level 2. Jones cites StrongDM’s three-person engineering team — which operates under explicit rules that code must not be written or reviewed by humans — and Anthropic, where 90% of Claude Code’s own codebase was written by Claude Code and project lead Boris Cherny hasn’t personally written code in months.

The core argument is that reaching Level 4 or 5 requires more than tool adoption — it demands end-to-end organizational redesign. Sprint planning, standups, Jira boards, and traditional code review all exist as responses to human limitations; when AI is doing the implementation, those structures become friction rather than coordination. Organizations seeing 25–30% or greater productivity gains, Jones argues, are those that went back to first principles and rebuilt their development workflows around AI capabilities — changing how they write specs, what they expect from different seniority levels, and how their CI/CD pipelines catch the new categories of errors that AI-generated code introduces.


📺 Source: Nate B Jones · Published February 18, 2026
🏷️ Format: Deep Dive

2 Items

Companies