Loading video...

Video Failed to Load

Go Home

Anthropic engineers just showed how they build a full app from scratch, using a loop of agents 40 minutes from the team behind Claude Code they used three agents: one to plan, one to build, one to judge, cycling until the app actually works the winners won't have the...

1,936,099 views • 4 months ago •via X (Twitter)

39 Comments

Bonsai 🌳's profile picture
Bonsai 🌳4 months ago

People are starting to learn more about Loop, looks like we’re not early anymore

Anatoli Kopadze's profile picture
Anatoli Kopadze4 months ago

I think the most important thing is not to be too late.

Brohonatron's profile picture
Brohonatron4 months ago

I look at this, and several of these recent Anthropic videos, and all I can think about is increased compute power, which equals money. So, you're doing something quicker, but the AI overhead cost is more than a human.

CyrilXBT's profile picture
CyrilXBT4 months ago

W share bro

Anatoli Kopadze's profile picture
Anatoli Kopadze4 months ago

appreciated!

echoes2099's profile picture
echoes20994 months ago

majority of peeps/companies can't afford to run this

Anatoli Kopadze's profile picture
Anatoli Kopadze4 months ago

That’s right

Khairallah AL-Awady's profile picture
Khairallah AL-Awady4 months ago

Great share my man! Bookmarked it

Anatoli Kopadze's profile picture
Anatoli Kopadze4 months ago

Appreciated

Crazyox🌶️(微微辣版)'s profile picture
Crazyox🌶️(微微辣版)4 months ago

三个agent循环出活,这思路真的香,回去也想试试搭一个

Rahul's profile picture
Rahul4 months ago

Loops or goal Use it and be amazed 😀

Ruth Heasman 🌳🌷🦚🐉😃's profile picture
Ruth Heasman 🌳🌷🦚🐉😃4 months ago

@Replit already does this.

Marv's profile picture
Marv4 months ago

the judge is the real engineering problem here. plan and build can be noisy and self-correct. but the judge has to know what "done" looks like before the loop even starts. that exit condition is harder to write than the app itself.

Brook Zimmatore's profile picture
Brook Zimmatore4 months ago

@grok analyze the entire agentic model in the video transcript and give an estimate of how many tokens it would cost and total $ amount estimated to do this

Cogito Ergo Sum's profile picture
Cogito Ergo Sum4 months ago

This actually is not very good representation of Loop in the Video. They could have done better

Rob's profile picture
Rob4 months ago

Retard

Kong Trading 🦍's profile picture
Kong Trading 🦍4 months ago

Forty minutes for a full app is terrifyingly fast

/var/log's profile picture
/var/log4 months ago

Learning the concept. Thanks

Springblade 🇺🇸's profile picture
Springblade 🇺🇸4 months ago

Ha, I just did this exact same thing myself yesterday morning, a 4-level, 3 Developers per level perfect coding machine! and by the end of the day my Dev Loop machine and Hermes and I had a fully built and working grow room control app, with UI Dashboard and everything sending sensor readings to the database, without one single error!! Grok made me this summary: CodeSquatch is my personal AI-powered "dev team in a box." Instead of one AI trying to build software alone (which often makes mistakes), it acts like a full company development department with 4 specialized layers: Junior Devs write the first version of the code. Senior Devs review and polish it for quality. Reg Devs (Regional Developers) make sure everything fits together properly, like a manager overseeing the big picture. Master Dev does a final hardcore check, fixes anything missed, adds smart improvements I didn't even think of, and runs tests until it's rock-solid. How it works: I give it a clear description of what I want. It first creates a detailed "Master Plan" for me to approve. Only then does the team start coding in stages, with lots of peer reviews and double-checks at every step. The goal is high-quality, reliable code with far fewer errors — basically a smart, tireless team that gets better every time I use it. It's completely reusable for any project (websites, apps, tools, etc.), runs privately on my local AI setup, and saves me huge amounts of time and frustration. It's my secret weapon for building stuff faster and better. (We just finished setting it up as its own independent tool/repo so I can improve it over time and use it everywhere.) SWEET!!🤌🇺🇸

Josh Jones's profile picture
Josh Jones4 months ago

Imagine that, Anthropic wants you to build loops that continuously use tokens. Weird

Danny Axelgod's profile picture
Danny Axelgod4 months ago

easy when you got unlimited tokens

Onur Eken's profile picture
Onur Eken4 months ago

a.k.a. burn tokens while you sleep do realize that tokens became a form of commodity, like electricity. no tokens -> no progress. it's the new abstraction of energy. which signals we're heading to a new "Tokendollar" economy. as most frontier models are US-built and proprietary. that's exactly why SOTA OSS models matter more then ever now -> fair access to tokens for everyone.

THC Humor 💹🧲's profile picture
THC Humor 💹🧲4 months ago

Agent collaboration might be a bigger unlock than better models

Andrew G. Huff's profile picture
Andrew G. Huff4 months ago

Loops? Hmm.

Kausik's profile picture
Kausik4 months ago

New way discovered to burn more tokens, make more money, great job @AnthropicAI to keep making Software Engineering to a Bullshit Engineering. Next up, how to keep running agents forever....coming soon.

Dreamweaver's profile picture
Dreamweaver4 months ago

Yes, let’s encourage people to leave high level product and critical engineering judgement to an agent that will spin its wheels using tokens till it gets it right over spending an extra 10-20mins to plan a high level product design spec with explicit logic instructions.

Steven Cheng's profile picture
Steven Cheng4 months ago

the loop is the real product

James AI's profile picture
James AI4 months ago

AI agents in a feedback loop will outperform AI models in isolation.

Hussain Hashim | Building SundayBack's profile picture
Hussain Hashim | Building SundayBack4 months ago

@AnatoliKopadze that's wild, never thought of using agents in a loop like that. makes me wonder how many iterations it took to get something decent though

Gregor's profile picture
Gregor4 months ago

not sure the judge agent holds in practice when i tried a similar loop building pennywise the judge kept greenlighting builds that broke on real data. what did they use as the termination condition?

Jonna L. Ward's profile picture
Jonna L. Ward4 months ago

The judge agent is where the cost hides. Caching the plan + build context with prompt caching cut my own Claude loops by ~60% on input tokens, since each judge pass re-reads the same spec.

Voltage (Fella)'s profile picture
Voltage (Fella)4 months ago

The plan→build→judge loop is a brute-force pattern that compensates for weak agency by adding more agents. If the builder can't self-evaluate, bolt on a judge. If the judge rejects, cycle back. You converge eventually, but you burn 3x the context per cycle minimum. Meh.

₥AN$A's profile picture
₥AN$A4 months ago

This is the way! Agentic workflows are becoming essential. My team at GroupGPT is building frameworks around similar loops. Btw, if you're into referral programs, ours pays 25% commission:

Ray's profile picture
Ray4 months ago

Plan → Build → Judge → Repeat — the smarter this iteration loop gets, the more reliable the output becomes. Thanks to share an important information.

Narasimhan Rengan's profile picture
Narasimhan Rengan4 months ago

Full video:

Cyberw@t's profile picture
Cyberw@t4 months ago

One agent plans. One agent builds. The other agent loops your tokens until you max out your credits

Pete McCormick's profile picture
Pete McCormick4 months ago

This is a fantastic share. Earned a follow :)

Alexander Menges's profile picture
Alexander Menges4 months ago

What would be cool: real demo on how to use it in operation. I think loops and workflows etc is pretty clear, but how to actually set them up (event-based loop triggers, multi-agent systems etc) would be helpful :)

DimiX's profile picture
DimiX4 months ago

this is exactly where coding is going. it’s no longer about individual prompt engineering, it's entirely about setting up the tightest agent loops

Related Videos