What are Agentic Loops?
Professor Ross Mike explains agentic loops — AI systems that self-iterate without human input — and argues they are largely impractical for most developers due to high token costs and poor assumption-making. He distinguishes between 'human in the loop' (iterative, guided development) and fully automated loops, recommending loops only for constrained, binary-feedback processes like code review. He shares his own GP loop use case as a rare valid example.
Summary
The episode features Ross Mike explaining agentic loops to host Greg. Ross begins by defining the traditional 'human in the loop' workflow, where a developer prompts an AI agent, reviews the result, and iterates manually — a cycle that keeps humans in control of direction and decisions.
He then contrasts this with agentic loops, where the AI generates a result, feeds that result back into itself as input, and continues working autonomously without human intervention. This approach is popularized by high-profile AI figures like Boris and Peter, who claim to build loops rather than write prompts. Ross argues these individuals have virtually unlimited token budgets (Peter reportedly burned $1.3 million in tokens in one month), making their workflow irrelevant as advice for average developers on $20–$200/month plans.
Ross's core criticism is twofold: first, agentic loops burn enormous amounts of tokens; second, no planning document can fully capture a developer's product vision, meaning the AI will inevitably make assumptions that misalign with what the builder actually wants. He compares it to hiring a developer, giving them a spec, and having them build the entire product without any check-ins — the result will be full of misaligned decisions.
He does, however, share a legitimate use case: his own 'GP loop' for code review. After pushing code to GitHub, a code review agent (Gravile) scores the code out of five. If the score is below four, cursor reads the review, makes fixes, pushes again, and repeats until a score of four or five is achieved or five attempts are made. He argues this works because the feedback loop is fixed, binary, and clearly defined — not open-ended or creative.
Ross and Greg conclude that loops are appropriate when output is binary (e.g., code review, SEO page generation) but inappropriate for creative, evolving product development. Both acknowledge that agentic loops may become viable in the future but argue they are not ready for mainstream use as of the recording date.
Key Insights
- Ross Mike argues that agentic loops are impractical for most developers because no planning document can fully capture a product vision — the AI will inevitably make assumptions that diverge from what the builder actually wants, analogous to a developer building an entire product without a single check-in.
- Ross Mike points out that advocates of agentic loops like Boris and Peter operate with essentially unlimited token budgets, with Peter reportedly burning $1.3 million worth of tokens in a single month — making their workflows financially irrelevant as advice for developers on standard subscription plans.
- Ross Mike describes his 'GP loop' for code review as a rare valid use of agentic loops: cursor reads a Gravile review from GitHub, makes fixes, pushes code, and repeats until the code scores four or five out of five — stopping after five attempts. He argues this works because the feedback is fixed and binary, not open-ended.
- Ross Mike observes that even his constrained code review loop breaks down when a push exceeds 1,000 lines of code, because the agent cannot fully contextualize and review that volume — illustrating that even well-designed loops have meaningful failure conditions.
- Ross Mike claims that agentic loops make sense only when the output is binary — such as code review scores or templated SEO pages — but are inappropriate for app development, which requires creativity, evolving decisions, and user feedback midway through the build process.
Topics
Full transcript available for MurmurCast members
Sign Up to Access