@AnnatarXBT: CLAUDE + LOOP + GRAPH ENGINEERING: ANDREW NG'S PLAYBOOK FOR AGENTS architecture beats model choice. every time it start…

CLAUDE + LOOP + GRAPH ENGINEERING: ANDREW NG'S PLAYBOOK FOR AGENTS architecture beats model choice. every time it starts with one loop. an agent revising its own work. everything else is that same idea, scaled up reflect -> tools -> plan -> collaborate -> graph > reflection - a critic checks the output against a rubric. 10-30% quality lift, day one > tools - code, search, a database. stops hallucinating, starts checking > planning - a plan gets written first. reroutes when a tool fails > multi-agent - coder, reviewer, tester. different rubrics catch different errors > graph - shared state everyone reads and writes to. the loop forgets, the graph doesn't what it looks like on disk: agent-system/ ├── loop/ │ ├── critic.md - the rubric it checks itself against │ └── http://revise.sh - reruns until it passes ├── tools/ │ └── tools.json - code exec, search, db query ├── plan/ │ └── plan.md - steps written before it acts ├── roles/ │ ├── coder.md │ ├── reviewer.md │ └── tester.md └── graph/ └── state.json - the memory every role reads and writes a loop externalizes revision. a graph externalizes state. each stage fixes what broke in the one before it ran this on Claude this week. changed how i structure every agent system now full breakdown below