Writing

Same words, same day. VYNN built a 10-tab model in 38 seconds and a 38-page report in 2 min 44 s. What was inside each, VYNN today, and how it got here.

I replaced VYNN's LangGraph pipeline with one general agent and 20 tools for stocks, funds, crypto and prediction markets, and made every number checkable.

Agent harnesses are operating systems built one feature at a time. This essay designs one as a whole, with git, not markdown, as its memory.

What a benchmark agent needed to work for developers: an IDE plugin, feedback it reruns from, a Self-Fix agent and a syntax-tree merge that keeps their edits.

The first VYNN AI, built alone: code computes every number, the model handles language. What held up, what I got wrong, and what a proactive VYNN needs.