Only the best are using them...

TL;DR

• 1

  Loops replace prompt-by-prompt coding: Instead of telling Claude or Cursor what to do each step, you define a trigger plus a goal, and the agent keeps iterating until it decides the goal has been met.

• 2

  A loop needs only two parts: Berman reduces the idea to a trigger and a verifiable end state, such as passing tests, green CI, or an LLM-based judgment for fuzzier product goals.

• 3

  The practical example is PR review and auto-fixing: In Cursor Automations, he sets a loop to run whenever a PR opens in AstroHub, review the code, fix issues, make tests pass, and commit changes back to the same PR.

• 4

  The hard part is not coding, it is specifying the goal: Deterministic targets like "all tests pass" are clean, but "build this feature" often requires a full spec up front, which clashes with how many builders actually explore and refine features as they go.

• 5

  Loop engineering is powerful but brutally expensive: Berman says only the "top 1% of 1%" are really using it today, pointing to Peter Steinberger's reported $1.3 million monthly token usage and the effectively unlimited budgets at OpenAI and Anthropic.

• 6

  Automation is not the same as a loop: An automation just runs a sequence, while a true loop contains a decision about whether the goal has been reached, which is why Berman connects the idea directly to reinforcement learning.