Video wird geladen...
Video konnte nicht geladen werden
anthropic engineers just built a full app from scratch with a loop of agents, start to finish in 40 minutes it's the claude code team. they used three agents: one plans, one builds, one checks the work. they keep cycling until the app actually runs the model matters less... show more
15,267 Aufrufe • vor 4 Tagen •via X (Twitter)
18 Kommentare

Great workshop, watched the whole thing and it was worth every minute

thanks dude

I like the plan / build / check split. Plan agent is the one I'd watch tho.. a bad brief just gets built faster now :)

lo que me parece más útil es que el que revisa no sea el mismo que escribe, si no se da el visto bueno a sí mismo

lo interesante es que siguen ciclando hasta que arranca, no se conforman con la primera versión

that's the part that makes it work, the first version is never the one that runs

saving this for later, nice content dude

appreciate it, worth the 40 minutes

the 40-minute end-to-end is interesting. how did they handle the agent failures - retry with same prompt or route to a different agent?

lo de que el modelo importa menos que el bucle es la frase que más gente debería leer hoy

me lo guardo, justo estoy montando algo parecido y lo tengo todo en un único agente

I’d love to see it when you’re finished with it

que lo enseñe el propio equipo que construye la herramienta le quita toda la duda

hard to argue with the team that builds the thing

The loop had distinct jobs: one agent planned, one built, and one checked whether the app actually ran. That validation step is what closes the gap between generated code and a working app.

the checker agent is doing the heavy lifting here. planner and builder are easy, getting something that actually says "no, redo it" and is right about it is the hard part

@aiscwork barbaridad de vídeo, todo el mundo debería verlo

A planner, a builder, and a checker looping until the app actually runs is a neat pattern, and it fits the idea that the harness around the model matters as much as the model itself. You and @mlkbfrey are the two accounts I enjoy following most.

