WTF Is an "AI Agent Loop"? Genius or Hype?
Greg Isenberg ·2026-06-09 ·1 min read
Summary written by us from the video's transcript. The video, and everything in it, is Greg Isenberg's work.
Learn what an AI agentic loop is, why it’s hyped, its practical limits, and when it can actually be useful.
Takeaways
- Agentic loops replace the human after the first prompt, feeding AI‑generated results back into the model.
- Unlimited token usage can make loops financially unsustainable for most developers.
- AI will inevitably make wrong assumptions when left to fill unspecified details.
- Loops succeed in constrained, binary‑feedback scenarios like automated code review or bulk SEO page creation.
- A practical loop can be built by tying a code‑review AI’s score to an iterative fix‑and‑push cycle in Cursor.
What an Agentic Loop Looks Like
In a typical *human‑in‑the‑loop* workflow you prompt an AI (Claude, Cursor, etc.), get a result, review it, and then manually iterate. An *agentic loop* removes the human after the first step: the AI receives its own output as new input, evaluates it, and continues generating until a goal is met.
Why the Hype?
Leaders like Boris and Peter promote loops because they promise fully autonomous development—feed a spec.md file to an agent and let it build an entire product without further human guidance. The idea sounds futuristic and could accelerate prototyping.
The Core Problem: Assumptions and Cost
When the AI runs unchecked, it must fill gaps in the specification on its own, often making wrong assumptions about design, architecture, or user experience. Those mistakes cost time and, more critically, burn large numbers of tokens—potentially millions of dollars for unlimited‑budget users.
When Loops Actually Work
Loops are viable in tightly bounded tasks with clear, binary feedback. Examples from the conversation include: *code review* (using Greptile or Code Rabbit) where a score out of five drives a repeat‑until‑acceptable loop; and large‑scale SEO page generation where uniform output is acceptable.
A Real‑World Loop Example
Ross describes his own setup: code is pushed to GitHub, an AI review agent assigns a quality score, then a Cursor “Grep loop” reads the review, makes fixes, pushes again, and repeats until the score reaches four or five. The loop stops after a set number of attempts or when the target score is achieved.
Practical Advice
- Treat loops as experiments, not production pipelines for core products. - Keep token budgets in mind; unlimited spending only makes sense for research teams. - Use loops only where feedback is objective and repeatable (e.g., linting, SEO, code quality).