THOUGHTS

How much information does JEV need to make a useful decision, and when is a custom classifier worth training? An exploration of context, labels, and evaluation.

The agents could already code, test, and review. What they lacked was organization. Once roles, handoffs, evidence, and blockers persisted beyond individual sessions, my daily became a short review of the decisions that actually needed attention.

While building a coordination system for agents, I found a broader question: what if agent systems borrowed proven roles, governance, and institutional memory instead of inventing an organization from scratch?

Linux already ran on almost anything. The hard part was handling the exceptions. Omarchy, agents, and skills are making those exceptions cheaper to diagnose.

Vibe coding is everywhere lately—mostly as a meme or criticism aimed at people trying to build software without really knowing how to code. But… what happens when it's used by an experienced engineer solving a real problem?