WTF Is an "AI Agent Loop"? The truth.

TL;DR

• 1

  Most agentic loops are expensive hype for normal builders: Ross Mike says tools like /goal and similar loop commands can chew through token budgets fast, especially if you are not already on a $200 per month plan.

• 2

  The core problem is hidden assumptions: a giant PRD or spec.md never captures every product detail, so an agent left alone fills in the gaps itself, often in ways that drift from the founder's actual vision.

• 3

  Human in the loop still beats agent in the loop for startup building: for apps, SaaS, and websites that need taste, iteration, and user feedback, Ross argues you still need to steer feature by feature instead of pressing go and hoping.

• 4

  Loops work better when success is measurable and constrained: Ross and Greg both point to binary tasks like code review, SEO page generation, and other fixed-output jobs as the places where loops make the most sense right now.

• 5

  Ross's real loop is a code-review loop with Gravile and Cursor: he pushes AI-generated code to GitHub, lets Gravile review it, then runs a "gp loop" skill that reads the review, fixes issues, and repushes until the score hits 5 out of 5 or five turns pass.

• 6

  Even the good loop breaks under scale: when a pull request exceeds about 1,000 lines, Ross says his review loop often fails because the agent cannot fully understand the whole diff, so he has to split the work into smaller PRs.