Plan, build, judge: why the agent loop beats the smarter model

Engineers from the team behind Claude Code published a 40-minute video building a full app from scratch, and the structure is the entire point: three agents in a loop. One plans, one builds, one judges, cycling until the app actually works. The clip pulled 1.8M views, and the line everyone quoted was "the winners won't have the smartest model, they'll have the best loop." ^[1]

I think that line is correct. I also think it's the most over-sold sentence in AI this month. Both can be true.

What's actually new

Nothing about plan, build, judge is novel. It's the same generate-and-verify pattern people have hand-rolled since the first coding agents shipped. What changed is who's demoing it and how cheap each iteration got.

The load-bearing agent is the judge. A planner and a builder, on their own, just give you a faster way to produce plausible-looking code. The judge is the agent that runs the thing, reads the failure, and sends it back. That's what turns a one-shot guess into a loop that converges on working software. Take the judge out and you have autocomplete with extra steps.

This is why "best loop beats smartest model" lands. A weaker model inside a tight verification loop will out-ship a frontier model answering in a single pass, because the loop catches the mistakes the model can't see in one shot. The intelligence moves from the weights into the control flow.

The part to be skeptical of

The source is a tweet thread, not a whitepaper and not a public repo. "Full app in 40 minutes" is the kind of framing that survives exactly until you ask what the app was. I'd hold the time claim loosely until someone publishes a build you can reproduce. The pattern is real. The benchmark is a vibe.

So I'm separating two claims that the viral post fuses together. The agent loop as an architecture is worth copying today. The specific 40-minute result is a marketing artifact, and I'm not citing it as a fact about your project.

The evidence that actually holds

The durable case for loops isn't a staged clip, it's adoption. Spotify's chief architect said more than 99% of their engineers now use AI coding tools, with uptake accelerating after Opus 4.5, at an org shipping roughly 4,500 deployments a day. ^[2] That's a production company reorganizing around AI-assisted shipping, not a 40-minute demo.

The market this points at is bigger than tech. Mark Cuban argued the real AI wealth transfer goes to the 33 million companies that "aren't going to have AI budgets" or AI experts, the shoe store and the regional trucking outfit and the small accounting firm. ^[3] A repeatable plan, build, judge loop is precisely the thing that lets one builder serve those businesses without standing up a research team. The loop is the leverage.

Where the loop breaks

A loop is only as honest as its judge. If the judge checks "does this look right," it will happily approve confident garbage, and you'll burn tokens converging on something broken. The judge needs an objective signal: a passing test suite, a server that returns 200, a screenshot that matches. Give it prose and it grades on vibes.

The second failure mode is no exit condition. A loop that can always find one more thing to fix will keep going until it talks itself into a regression. Cap the iterations and make "done" a defined state, not a feeling the model reports.

If you're wiring this up today

1. Build the judge first. Your loop is only as good as the agent deciding "not done yet," so design that agent before the builder.
2. Give the judge something objective to check, like tests, a running process, or a diff, instead of asking it whether the output looks correct.
3. Cap the iterations and define an explicit success state, so the loop terminates instead of grinding.
4. Don't wait for an official template. The agent loop is public enough to build this week, and the people shipping with it already are.

The model you pick matters less than most of the timeline thinks. The agent loop you build around it is the part that compounds.

Sources

Primary source: Anatoli Kopadze (@AnatoliKopadze) on X

1. Tweet by @AnatoliKopadze — @AnatoliKopadze · community
2. Tweet by @0xMovez — @0xMovez · community
3. Tweet by @r0ck3t23 — @r0ck3t23 · community