Making $$$ with Loop Engineering

Loop engineering is a build-verify-iterate cycle for AI agents consisting of a build step, a verify step, and a stop condition. The concept traces back to lean startup and lean manufacturing principles from Toyota, popularized roughly 10 to 15 years ago. Boris from Claude Code and Peter Steinberger from OpenClaw brought the term into wider circulation on Twitter approximately one month before this episode. The core mechanic is illustrated by Claude Code's slash goal command, where a secondary agent checks whether the main builder agent has finished and loops until the goal is achieved, with a stop condition required to prevent infinite iteration.

Ellie runs a production SEO loop for Inbox Zero, an AI email management product, and for draftfantasy.com, a site he started approximately 12 years ago. The loop connects to Google Search Console and the Data for SEO API to retrieve real ranking data and competitor comparisons, makes page-level changes, records actions in a markdown file, and runs once every two weeks to one month. Inbox Zero currently ranks first on Google for the search term inbox zero and approximately position 30 for AI email assistant, with a domain rating of roughly 63 to 64. Draftfantasy.com ranks at an average position of 4.4 for a key term, generating roughly 120,000 clicks and approximately one million impressions over three months, with an estimated potential of around half a million clicks if rankings improve to position two or three. Early results show some terms moving from page three to page two, and SEO gains can compound suddenly after a period of little visible progress. SEO improvements also benefit LLM and GEO search ranking, not only Google.

The verify step does not need to be product-related; any objective metric serves as the feedback input. For Inbox Zero, the stop condition is email categorization accuracy, with the agent retrying until it reaches a 90 percent accuracy threshold. For an SEO loop the objective metric is ranking position. For a Facebook ad loop the objective would be profitability, with the AI generating ad variants, monitoring performance, and shifting budget toward winning variants. AI is well-suited to testing lines of copy and text-based formats like Google search ads, but AI-generated video and graphic creative is likely inferior to human work, making a combination of human creative input and AI optimization the recommended approach.

The cost of running an SEO loop once per month is estimated at under five dollars in API tokens, described as negligible compared to hiring an SEO agency or freelancer. The open question is whether agents are currently capable of matching or exceeding the performance of a human agency. Loop costs can reach 50 to 200 dollars per month at higher usage levels, so adding a stopping condition tied to business value, such as assigning a monetary value to a click or a customer, is recommended to cap spending. Peter Steinberger was cited as spending 1.3 million dollars per month on AI credits. Users on a 20 dollar plan should be cautious about token consumption and consider open source models, while users on a 100 to 200 dollar per month Max plan receive large token allocations. Setting up Slack ping notifications each time a loop run completes allows review and approval of results before changes go live.

A product feedback loop would have an AI agent read customer feedback, analyze tools like PostHog, review logs and Sentry error reports, prioritize pain points, prototype features, and fix bugs. Separating a bug loop focused on uptime from a product feedback loop focused on metrics like DAU, MAU, retention, or virality is recommended. A fully autonomous product feedback loop running on a live business would carry significant risk. Dmitro tweeted in 2026 that software should be able to build itself and achieve product market fit autonomously with the founder only needing to fund tokens, though Ellie notes this was intended as satire of loop engineering hype. Companies attempting to build businesses almost entirely through autonomous AI loops are expected to emerge within the next year, with early experiments already occurring but without exceptional results so far.

For social media, performance signals such as likes, impressions, and conversions can be fed back into the AI so it iterates and improves over time. The preferred metric for a social media loop is average views per week, with the goal of pushing that number up continuously. Ellie is skeptical that an AI loop could reliably grow a Twitter account to 100,000 followers. A more practical starting point is a minimal viable loop optimizing for a verifiable threshold such as 10 likes per post rather than targeting a large follower count. Ellie argues that every part of a business, including social media, video content, cold outreach, and customer support, could potentially be placed on a loop, with AI functioning like a founder who wakes up each day asking how to improve operations.