A calm place for ideas.https://wuwei.page/https://wuwei.page/en/blog/how-the-agent-loop-works/https://wuwei.page/en/blog/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.Tue, 07 Jul 2026 00:00:00 GMThttps://wuwei.page/en/blog/ponytail-less-code-measured/https://wuwei.page/en/blog/ponytail-less-code-measured/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.Tue, 07 Jul 2026 00:00:00 GMThttps://wuwei.page/en/blog/cutting-noise-without-cutting-signal/https://wuwei.page/en/blog/cutting-noise-without-cutting-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.Sun, 05 Jul 2026 00:00:00 GMT