Last 30 Days: The Loop

The 30-day sweep behind today's Run Data Run post: the moat thesis, the skeptics who got there first, the cost numbers nobody pitches, and the one stat I refused to print

About Last 30 Days. Cross-platform research sweeps on topics worth paying attention to. Every post pulls Reddit, X, YouTube, Hacker News, Polymarket, and the web from the last 30 days, then synthesizes what people are actually saying, building, and betting on. Topics get picked when the signal is high and the story is contradictory, when a single headline would lie about the shape of what's happening. Each post follows the same arc: one specific finding that earns the click, why the topic deserves a sweep right now, the themed synthesis with inline citations, and the follow-up threads worth watching next.

Today's Run Data Run post, The Loop Is Simpler Than It Sounds, rests on a 30-day sweep across every platform where people argue about AI. This is the companion: the raw shape of what came back, including the parts that did not fit the headline. If the post is the argument, this is the receipts.

The single most useful thing the sweep found was not a fact. It was a contradiction. The loudest claim of the month, that "loop engineering" is the new moat, sits directly on top of an older, quieter body of evidence saying the exact opposite: that loops fail in ways the hype never mentions, and the people who learned that learned it expensively. Both are real. A single headline would have to pick one and lie about the other.

Why this topic deserves a sweep

In June, "the loop" went from a clever bash trick to the most over-discussed idea in the field. Anthropic shipped it as a product feature. A conference talk produced a slogan, "the winners will not have the smartest model, they will have the best loop," that got repeated everywhere. Boris Cherny, who built Claude Code, said he has not hand-written code in eight months and now spends his time writing loops, some days managing thousands of agents.

That is one half of the story, and if you only read X you would think it was the whole thing. The sweep is what surfaces the other half: a counter-current nearly as strong, mostly predating the viral wave, made of cost blowups, reliability math, and production failures. You cannot read both halves and conclude one thing. That is exactly when a sweep beats a headline.

The moat thesis, as it actually spread

The viral layer traces to one talk. The highest-engagement post of the set put the slogan plainly (@AnatoliKopadze, 8,624 likes, 1.7M views), describing Anthropic running "three agents: one to plan, one to build, one to judge, cycling until the app actually works." The clearest single statement of the mechanic came from a practitioner, Akshay Pachaar: "the loop itself is six lines, and nobody competes on it. every serious agent framework lands on the same tiny while-loop."

The technique has a name and an origin: the "Ralph" loop, coined by Geoffrey Huntley after Ralph Wiggum, because it looks too dumb to work. And it has a button now, Claude Code's native `/loop` and `/goal`/loop and /goal](https://code.claude.com/docs/en/goal) commands, which is usually the signal that something has gone from clever to standard.

The one stat I refused to print

Here is a transparency note worth making, because it is the kind of thing a sweep catches and a single source does not.

The most-quoted number in the whole topic, Anthropic's internal loop-adoption figure, does not survive the sweep. Three different posts attribute three different numbers to what appears to be the same talk: "over 30% of code" written through loops, then "70 to 80% of engineers" using them, then "90%." When one stat arrives at three sizes from one source, it is folklore, not data. So the post states the direction (Anthropic is clearly building this way, at scale, on the record) and explicitly tells you to treat the decimal point as noise. That call only gets made because the sweep put the three versions side by side.

The counter-current the slogan skips

This is the densest part of what came back, and the part the hype leaves out.

• The math. Melanie Warrick at Temporal ran the numbers: even at 85% reliability per step, a ten-step workflow finishes correctly only about 20% of the time. Failures multiply. She found the gap is resilience, not intelligence; agents "diagnose a failure in perfect detail and still do nothing to recover from it."

• The cost. An arXiv study found the same task, same model, same prompt, could cost eight dollars or two hundred forty, thirty times apart with no input change. A separate $22,000 sweep pinned a 33-fold spread to scaffold choices. And one widely-shared account had Uber burning its annual AI coding budget in four months before capping engineers at $1,500 a head.

• The verification trap. A benchmark of agentic systems spent roughly $40,000 on evaluation runs alone, and noted that doing the checks rigorously would push it into the hundreds of thousands. If the whole premise of a loop is a verifiable done-condition, and verifying "done" at scale is the single most expensive line, that is a price nobody put in the pitch.

• The adoption gap. By one survey, 79% of enterprises have adopted agents and only 11% run them in production. Gartner projects 40% of agentic AI projects will be canceled by 2027, on cost, value, and governance, not on the loop failing.

Gary Marcus and Arpit Bhayani carried the skeptic load on the social layer, both before the June wave crested. The reliability and cost objections were live before the moat framing peaked, which is itself a signal.

How the sweep is built (the method)

A quick word on method, since the point of this section is to show the work.

The sweep pulls each platform on its own terms: Hacker News and Reddit for builder substance in the comments, X for engagement-weighted reach, a curated corpus and a Substack named-voice pool for depth, and a separate open-ended discovery pass whose only instruction is "surprise me, do not confirm the thesis." That last pass is what surfaced the counter-current; left to its own framing, a research pass confirms what you already believe.

Two integrity habits matter. Every cited cost figure traces to a primary source, not a tweet about a source. And the fetch pass flags its own failures: of the URLs pulled this round, several came back as bot-challenge pages or 404s and were routed around rather than quoted. A claim that only survives in one blocked link is not a claim.

What I'm watching next

• Whether the Anthropic adoption number ever gets pinned to a primary transcript, or stays folklore.

• Where exactly the "coherence cliff" bites, by task length, token count, or iteration count. Everyone asserts it; nobody has located it.

• Whether the three-agent plan/build/judge pattern survives the move from greenfield demos to brownfield maintenance, where the stale-state failures are worse.

• The shift the researchers are naming: the agent stops being a tool that runs your code and becomes the software, with the human as "intent architect." If that holds, the unit of cost moves from per-token to per-finished-artifact, and most budgeting models break.

Sources and method

Full synthesis and the per-platform raw dumps live in my research vault; the post, The Loop Is Simpler Than It Sounds, is the argument this sweep grounds. Method: a 30-day cross-platform sweep (Reddit, X, YouTube, Hacker News, the web, a curated corpus, a named-voice pool) plus an open-ended discovery pass, two-pass synthesis, every quantitative claim traced to a primary source. Counts and quotes come from the actual sweep output, not memory.

---

Last 30 Days is a research series on Run Data Run, posted alongside the occasional Deep Dive when a topic earns the deeper look. No email on these, they live on the site for when you want the receipts behind an argument. If a sweep was useful, subscribe and you'll get the next one as it lands.