How coding agents improve through language (specs, reviews, lessons, rules) without touching model weights.
| Page | About |
|---|---|
| index | Thesis, three layers, evolution from vibe coding to meta-programming |
| specification | DO/DON'T/GLOSSARY, SDD, Intent Formalization, AGENTS.md |
| context-engineering | Token budgets, progressive disclosure, caching |
| pipeline | Scout → Spec → Plan → Workers → Review → Lessons |
| verification | Separate reviewer, multi-model, OTel, Amazon case study |
| self-improvement | Memory hierarchy, lesson extraction, closed loop |
| principles | 6 principles, 3 maturity levels, anti-patterns |
| playbook | 15 rules tied to principles and experiments |
| landscape | Tools, trends, papers — April 2026 snapshot |
| references | Full bibliography — papers, repos, people, our experiments |
| changelog | What's new since previous sync |
- 🟢 Our experiment, our data
- 🟡 Trusted source (Anthropic, Microsoft Research, peer-reviewed)
- 🟠 Community reports (practitioners, GitHub projects)
- 🔴 Unverified
- ⚪ Our opinion / synthesis