Agentic Notes
Building production AI workflows with Claude Code, and the method behind it, written down so it is repeatable.
Loop Engineering
How an AI Team Builds an App End-to-End. When agents build for almost nothing, the only critical path left is where you wait for a human and where the agent guesses. Loop engineering batches all the waiting to one point up front, kills every guess, and lets a team of agents run the sprint loop themselves — with the receipts behind every rule, and an honest note on the part not yet through fire.
The Agent Gym
Where Skill-and-Tool Training Actually Runs. The prior essay argued that building an agent IS training it. This one is the room where that training runs for real — a gym with a teacher and a courtroom. The loop, the diagnosis tree that keeps you from blaming the student first, and the discipline that stops the whole thing from lying to you. With receipts from a system that graduated.
Bootstrap một đội AI build app
Build-System Playbook — toàn văn để dựng lại. Bộ khởi động để một session Claude Code tự dựng một đội AI (architect · backend · frontend · tester) và build app end-to-end trong vòng lặp sprint tự trị. Bài này đăng NGUYÊN VĂN từng artifact — đọc một URL là dựng lại được cả quy trình.
Agent-Building Is Training
And You Hold the Key. Validate picks the starting point. To make an agent good for YOUR problem you have to train it — tool and skill are the weights, the scorer is the loss. Here is that loop, and the information-theoretic reason you (the human) are the one weight that cannot be removed.
Validate AI Workflows 50x Faster
No-Build-First Method. A tool-agnostic methodology to separate experimentation runtime from production runtime — saving weeks of engineering before writing a single line of production code.