A deep dive into the "design-by-intuition" tax in AI engineering, comparing convergent and divergent agent loops. It introduces agent-evolutions, a genetic search framework that spawns variants in parallel, evaluates them against frozen binary gates and numeric rubrics, and relies on a robust file-based sub-agent contract rather than chat parsing.

Many coding agents fail because they improvise internal planning loops, leading to unchecked code or runaway sessions. This article details how externalizing a structured Plan-Do-Check-Act (PDCA) engineering loop into a disk-based workflow acts as a leverage multiplier. By forcing agents through an explicit, step-by-step verification process, developers can reliably guide agents to finish what they start with minimal human intervention.

In this episode, we're joined by Josh Collier, FinOps Lead at Superhuman (formerly Grammarly), to explore what it really costs to run AI at scale and why the rules of the game changed faster than anyone expected.