Descriptions:
Greg Isenberg sits down with builder and educator Ross Mike to cut through the hype around agentic loops — the fully autonomous AI development cycles that prominent figures like Boris and Peter have claimed to use instead of writing prompts manually. The conversation draws a precise distinction between human-in-the-loop workflows, where developers actively direct and review AI outputs at each step, and fully autonomous loops where an agent checks its own results and continues working without human oversight.
Mike’s core argument is that agentic loops are impractical for most real-world builders. The fundamental problem is specification: it’s nearly impossible to fully encode a product vision into a single document, and the token burn from long autonomous runs can be catastrophic. He compares the experience to a slot machine — exciting, expensive, and largely unpredictable. Tools mentioned include Cursor, Claude, and Codex as common human-in-the-loop harnesses.
The episode does endorse one clear use case for agentic loops: low-stakes experimentation where implementation details don’t matter. Mike demonstrates with an Among Us-style AI benchmark simulator he built autonomously — a project where core functionality mattered but fit-and-finish did not. The takeaway is nuanced rather than dismissive: loops have a real place, but for meaningful production work, the human still needs to remain in the loop.
📺 Source: Greg Isenberg · Published June 09, 2026
🏷️ Format: Podcast
