Summary
Nate Herk walks through the concept of agent loops — a shift in how developers and power users should think about working with AI coding assistants. Rather than manually prompting tools like Claude Code, the argument is that serious practitioners now design loops that do the prompting for them. Industry figures Boris Churnney and Peter Steinberger are cited as examples of builders who have publicly moved to this model.
The video breaks down the three core components of a loop: a trigger, an action, and a stop condition. Herk explores three common architectures — a single reasoning-acting-observing agent, a maker-checker pair, and a manager with multiple helper sub-agents — arguing that the right structure depends on the use case rather than on running the most sophisticated setup possible.
A live demo shows an agent loop that researched 45 sources (articles, YouTube transcripts, social posts) and iterated through seven versions of an HTML artifact before completion, with Claude Code spending 27 minutes on thumbnail concept generation, scoring candidates against objective criteria, and selecting a winner through iterative refinement. Herk also references Matthew Berman’s public loop library as a starting resource. The central message: understand your goal and stop condition before scaling complexity, or you will simply scale your problems.
📺 Source: Nate Herk | AI Automation · Published June 19, 2026
🏷️ Format: Tutorial Demo
