What are Agentic Loops?

Ross Mike argues that most people discussing agentic loops, including prominent voices in the AI industry, fundamentally misunderstand what they are and how to use them responsibly. He draws a clear distinction: human in the loop means the agent builds while the human directs, governs, and approves each step, whereas an agentic loop removes the human entirely so the agent checks its own output and feeds it back into itself automatically.

Ross Mike singles out AI figures Boris and Peter as examples of people who advocate for building loops over writing prompts, but notes that Peter burned 1.3 million dollars worth of tokens in a single month by his own account. His core criticism is that Boris and Peter operate without token budgets and can burn unlimited tokens, which is why agentic loops work at their scale but are inappropriate for average users on 20 or 100 dollar per month plans. He sets the minimum threshold for even considering agentic loops at the 200 dollar per month plan.

When an agent is given the floor to make assumptions without human direction, Ross Mike says it gets things wrong most of the time. He illustrates this with an Among Us simulator built using an agentic loop in approximately one and a half hours that contained many incorrect details not specified in the initial instruction set. He also notes that a plan or PRD markdown file always has edge cases and missed details no matter how thorough it appears, and that it is impossible for a human to fully contextualize every product detail in a single document because requirements evolve.

Ross Mike considers agentic loops useful only for prototyping and experimenting with small tools where details do not matter, and calls them a catastrophe when product quality matters and token budgets are limited. He argues loops are only appropriate when the output is binary, meaning tasks with no creativity involved, and cites code review tools like CodeRabbit and Greptile and formulaic SEO page generation as legitimate use cases. He explicitly says loops do not make sense for app building, startup founders, or any open-ended creative goal.

Despite his general criticism, Ross Mike uses one loop himself, which he calls Greptloop. He installs Greptile as a code review agent on GitHub to review AI-generated code pushed from Cursor, where Greptile provides a score out of five and flags security issues, broken edge cases, and missing elements. He will not push code to production unless the Greptile score exceeds four out of five. The Greptloop instructs the agent to read the GitHub review, fix the issues, push the changes, and repeat until the score reaches five out of five or five turns are exhausted. He describes this as a closed, goal-oriented loop with a fixed and defined feedback mechanism, which is what makes it appropriate. The loop breaks when a code push exceeds 1,000 lines because the agent cannot fully review and contextualize that volume, requiring pushes to stay under that limit or be split into multiple pull requests.

Ross Mike acknowledges that sophisticated loop builders like Boris and Peter likely use test suites, browser-use capability, and a meta harness around the agent to improve loop reliability, but argues this infrastructure is not accessible or practical for most users. He predicts agentic loops will eventually work as broadly promised but states they are not practical for most users as of June 9, 2025, and concedes he could look wrong within a month if the technology improves rapidly. His conclusion is that human in the loop remains the best loop for anyone building something meaningful today.