WTF Is an "AI Agent Loop"? Genius or Hype?

TL;DR

• 1

  Most agentic loops are expensive hype for builders: Ross Mike says letting an agent run from a giant PRD without human checkpoints leads to bad assumptions and heavy token burn, especially for people on $20 or $100 plans instead of $200-plus plans.

• 2

  The core problem is missing human judgment mid-build: When an AI agent builds an app end-to-end, it has to guess product details, architecture, and UX choices that are rarely fully captured in a single markdown spec.

• 3

  Ross compares open-ended loops to a slot machine: Tools like /goal and /loop can work for prototypes and experiments, but for meaningful products they often keep spending tokens while drifting away from the actual vision.

• 4

  A good loop needs fixed feedback, not vague creativity: Ross's working example uses Cursor plus Greptile, where the agent reads a code review, fixes issues, pushes again, and repeats until the score reaches at least 4 out of 5, ideally 5 out of 5.

• 5

  Even the useful loop breaks under scale: Ross says his code review loop struggles once a pull request exceeds 1,000 lines, which forces him to split work into smaller PRs so the review agent can actually reason about it.

• 6

  The future case is real, but Ross says the timing is wrong: Greg and Ross both agree self-healing app-building loops may arrive later, but as of June 9, 2026, the safer default is still human in the loop.