Stop Using Agentic Loops Wrong

Automation dream promises a grand gesture: one prompt, full autonomy. This vision of agentic loops stands in stark contrast to human-in-the-loop workflows. Human-in-the-loop means you prompt, the AI builds, you review, then prompt again, step-by-step, directing every move to ensure alignment with your vision, much like Professor Ras Mic guides an AI through building a to-do app, feature by feature.

Consider the challenge of describing a complex tattoo over the phone. That's precisely the experience of initiating a Wide AI-open agentic loop. You articulate a single, comprehensive `/goal` for the agent to pursue autonomously, but such a prompt inevitably leaves critical gaps and invites significant misinterpretation.

This leads to the central problem of autonomous assumption: agents fill in details you didn't explicitly specify. Without precise direction, an agent guesses how to handle 'which screen shows up after login' or 'what happens when a payment fails.' Such assumptions often drift from intent, creating a slop machine that churns out flawed results and burns tokens, a costly reality for most users compared to the unlimited budgets of Boris Cherny Cherny or Peter Steinberger Steinberger, who reportedly spent $1.3 million in a single month.

Wide AI AI's promise of full autonomy often overlooks the staggering token costs. Builders like Peter Steinberger Steinberger have publicly shared burning $1.3 million in tokens during a single month pursuing Wide AI-open agentic loops. This model, where an agent continuously generates, reviews, and refines its own output from a single `/goal` prompt, inherently leads to massive consumption.

Such an approach is only viable for those with virtually unlimited budgets, like the well-funded operations of Boris Cherny Cherny and Peter Steinberger Steinberger. For the vast majority of developers and startups on tiered plans—the $20 or $100 tiers—these meta-harnesses quickly become a "slop machine," devouring budgets far too fast. The constant guesswork makes them impractical for real-world constraints.

Think of it like hiring a brilliant developer, handing over a high-level spec, and walking away. The agent, much like that solo developer, fills critical gaps with assumptions—details like "which screen shows up after login" or "what happens on payment failure." These guesses inevitably drift from the original product vision, leading to wasted cycles, missed targets, and an unexpectedly massive token bill for work that ultimately misses the mark.

Wide AI-open agentic loops, as discussed, become token-burning "slop machines" due to their reliance on unchecked assumptions. The secret to successful loops lies in the opposite: fixed, defined feedback systems. Agents thrive when operating within clear, quantifiable scenarios—think unambiguous 'pass/fail' or a specific numeric target. Without this binary battleground, the agent simply guesses, leading to expensive, off-target iterations.

Professor Ras Mic’s daily code review loop perfectly illustrates this principle. Using Cursor as the harness, GitHub for source control, and Greptile as the intelligent review agent, his system autonomously chases a perfect 5/5 score. Mic's `grep loop` command directs the agent: read Greptile's review, apply fixes, then push the changes and repeat. This iterative process continues until the code hits 5/5 or completes five turns, ensuring production-ready code always scores above four out of five.

This tight feedback loop works because the agent isn't making subjective design choices. It's optimizing against a precise, measurable outcome. The agent knows exactly what "good" looks like—a specific score from Greptile—and can methodically work towards it without drifting into costly guesswork.

This binary principle extends beyond code. Agentic loops shine in any task with a clear, measurable success metric: - Generating templated SEO pages - Performing repetitive data cleaning against a fixed ruleset - Validating configurations against a schema

For a deeper dive into loop engineering, including ReAct patterns and future developments, explore Agentic Loops Explained: From ReAct to Loop Engineering (2026 Guide).

Complex, creative endeavors like full application development resist complete agentic autonomy. The entire product vision lives within a human's head, requiring constant, subjective feedback and iterative refinement that machines cannot yet fully grasp. This prevents the token nightmare of Wide AI-open loops making costly assumptions.

Imagine trusting full self-driving for a road trip from Miami to Charleston without a single check-in. You would inevitably need to pull over, course-correct, and re-evaluate the journey based on unforeseen conditions or evolving preferences. AI agents face similar challenges with open-ended tasks.

Professor Ras Mic consistently emphasizes this reality: the human-in-the-loop model remains the strongest setup for most builders today. This approach prevents autonomous guesswork from burning millions in tokens, as seen with some top builders pursuing full automation.

While the future undeniably points toward greater autonomy, the most powerful and reliable workflow for complex tasks still integrates human intelligence at every critical juncture. For now, the best loop keeps a human hand firmly guiding the overall vision and validating progress.

What is an agentic loop?

An agentic loop is an AI workflow where a human gives a single, high-level prompt, and the AI agent then generates, reviews its own work, and iterates repeatedly until the task is complete, without further human input.

What's the main problem with agentic loops?

The main problem is that a single prompt rarely covers all edge cases. The AI fills these gaps with assumptions, which can lead to incorrect outputs (a 'slop machine') and extremely high token costs.

When are agentic loops most effective?

They excel at tasks with binary or clearly defined feedback, like pass/fail scenarios. Code review, where an AI can score code and iterate until it hits a target score, is a perfect example.

Is 'human-in-the-loop' still the best approach?

Yes. For complex, nuanced tasks like app development that require subjective feedback, human-in-the-loop—where a human reviews and guides each step—remains the most effective and cost-efficient approach today.