Writing
Tools, things built, and what's changing in how we program. Newest first.
- How the agent loop works An AI agent is not magic: it is a model calling tools in a loop, with the observations flowing back into the context on every turn. The mechanics underneath — think, act, observe — and where it breaks.
- ponytail: less code, better code? ponytail promises your agent writes ~54% less code without losing safety. Its own benchmark is honest (rare). I tested it where they didn't reach: strong model, a genuinely large repo (Django, 522k LOC) and a long trace — 4 A/B tasks, measured with git diff.
- rtk: cutting noise without cutting the signal rtk promises to cut 60-90% of the tokens a coding assistant spends reading terminal output. I installed it, tested it at real scale, and checked every finding against its own docs, PRs, and issues.