the art of simplicity

The 'goodbye loops, hello graphs' posts made me wonder whether graph engineering replaces the orchestrator/planner/coder/reviewer setup I'd already written about. It doesn't. One says who does the work, the other says how control and state move between them.

AI coding agents have a groundhog-day problem: every session they re-read the same files to rebuild the same mental model, then forget it all when the session ends. Graphify extracts your codebase into a graph; Google's Open Knowledge Format stores it as portable, git-versioned agent memory. Two commands, one afternoon.

Most of us aren't at a clean 'level' with AI coding agents; we're wedged between two. A field guide to the rungs and what actually changes as you climb: from pasting snippets into a chat box to running isolated, reviewed, token-lean agents you delegate from Slack.

Two open-source tools both hit ~60k GitHub stars in five months by attacking the same problem from opposite ends: AI coding agents spend most of their tokens on plumbing, not thinking. RTK compresses the output; CodeGraph eliminates the discovery.

A quick comparative tour through how OpenAI, Anthropic, Google, Microsoft, Shopify, and Stripe each approach the same problem: keeping coding agents on the rails.