- Author
- Alex Ewerlöf
- Published
- Sep 26, 2026
- Kind
- Long opinion post on Substack, written as independent micro-posts with later updates
- Original
- blog.alexewerlof.com/p/coding-is-not-solved
- Subject
- Coding with agents
What it is
Alex Ewerlöf wrote this post as a reliability engineer who built with AI. The author defines coding and solved, then argues from accountability, risk and economics. The post also lists fallacies, good AI uses and possible new job lanes.
What it says
- Ewerlöf says solved means AI output needs no inspection. Coding is not there.
- Maintenance, reliability and security are the majority of software cost, so cheaper creation matters less.
- AI cannot be held accountable, so people who ship code must understand it.
- Reading code is optional for only four kinds of product, such as proofs of concept.
- LLMs succeed at code mainly through feedback loops, yet stay stochastic where code is deterministic.
- The post puts the field at least two breakthroughs from not needing to read code.
- It lists good uses of AI. It also lists cases to avoid.
What evidence it gives
The evidence is the author's experience, quoted social posts and links to bug reports. One test had an agent update five npm dependencies in 12 minutes and 72 steps. There is no controlled comparison. Many claims about vendors' motives are the author's theory.
Interests and limits
The author says coding brings no income, but sells paid subscriptions and a book. Images and screenshots in the post were not viewed. Only two reader comments were shown on the page. The post is opinion, and its claims about companies were not checked.