10 things I learned from burning myself out with AI coding agents

Abstraction: Practical limits of AI coding agents after 50 real projects

Key points:

• Author completed 50 projects using Claude Code (Claude Opus 4.5 via Claude Max), OpenAI Codex, and Google Gemini CLI over several months
• AI coding agents excel at flashy prototypes of simple applications and games but only by borrowing patterns from training data — like a 3D printer for code
• Creating durable production code, managing complex projects, or crafting truly novel solutions still requires human experience, patience, and skill
• Claude Code described as the most extensively used tool; author paid out of pocket for both Claude Max and OpenAI premium plans
• The experience is described as the most fun with a computer since learning BASIC on Apple II Plus — enthusiasm despite limitations
• Analogy: AI coding agents are like 3D printers: impressive output, but polished production work requires far more than pressing a button

Connections: Claude Code · Anthropic · Openai · Agentic Coding

Source: https://arstechnica.com/information-technology/2026/01/10-things-i-learned-from-burning-myself-out-with-ai-coding-agents/