Anthropic Claude Code engineer: "If you're watching Claude write code, you're the QA tester. That's not what you're paid for." In 37 minutes he lays out how to get your keyboard out of the hot path entirely. The shift is /loop. You tell Claude to wake up every 10... minutes and babysit your PRs, and it just does it, while you're nowhere near the laptop. Routines do the same in the cloud, so the work keeps running with your machine closed. He caps it with remote control: any session, on any surface, driven from your phone. Watch the full talk, then grab the setup below.show more

This is anthropic trying to get you even more dependent on their product.

lmaoo, good theory

The slightest imprecision in the initial trajectory can bring super off track if you go really far. He must witness some pretty weird stuff done by his AIs on a daily basis.

I'll believe him if he was paying for his own tokens

Fairly useless for generating feature ideas, choosing between them, making sure integrated features are coherent, and fine tuning as during development. Fable was ok at it as it has better “taste” than other models.

we find the visuals in this presentation highly stimulating

Nice concept!! Loop is goated!

genuinely what use cases is this for? 80% of the work is writing and iterating on a good spec which cannot be done with a loop

Why does everyone keep calling cronjobs “loops”?

Agent loops change everything now

lol fuckin stupid hyped ate. sleep while other (agents) work and have no idea what's going on. and thounsand agents running? how much does it cost though? useless post as always crasp

/loop gets rid of the whole babysitting-Claude problem

QA’ing code is not a bad idea.

We have to babysit our agents because LLMs aren't modular reasoning systems. they can tell you all the abstract layers yet the moment they try to implement they forget all the rules and layers. why? because they are linear

“sorry, i need to correct something i told you at the beginning of this chat….” - opus 4.8 i’d rather babysit while or before it is writing the code, rather than finding the blunder during “merge to main”

I would like the loop also if i dont have to babysit. Sometime it just spam over engineered code, sometime duplicate code and sometime assuming code..

another video about loops, very good

It's all about the harness. i can run loops for hours, adding a feature every cycle, without degradation. last cycle as good as the first. 8/8 every cycle until loop finished All on the cheapest openrouter model, with 0 harm At least 40x less tokens than opus 4.8 and the quality differential isn't what you would think

Why pretending to teach engineering when it’s all about marketing to promote new features that are designed to rely so much on the agents and leave them unsupervised resulting in disastrous releases in production…

getting Claude to babysit PRs on a loop is exactly how you stop the "watching it type" trap. Remote control from the phone is gonna be so nice too

useful 37 mins

Stop prompting. Start orchestrating. 🚀

/loop is powerful; I wanted the status layer around it: which Claude/Codex run is running / waiting / stuck? Built Agent Island for the MacBook notch + auto-resume for selected long runs.

Exactly. The real win is not faster typing , it’s better intent, constraints, and quality control.

/loop is underrated. The win is not just fewer keystrokes, it's fewer context switches. A 10-minute wake cycle gives the agent time to explore, fail, and come back with options instead of turning the human into a full-time spinner loader.

Everyone should do this

Sounds like a great way to loop into chaos.

the /loop framing is right but there's a trap. i let agents run unattended on a refactor last month, came back to 40 files changed with 3 broken assumptions baked in deep. keyboard leaves the hot path, but defining 'done' precisely becomes the whole job.

This is not a silver bullet for QA. In large codebases agents miss one of the main causes of AI slop which is duplicated code (functions doing the same thing under different names). Probabilistic QA is not QA.

What is the most reliable way to monitor cloud routines?

i set up almost exactly this /loop wake-every-10-min thing and the part nobody warns you about is how fast it eats the weekly quota. only saw it coming after i started watching the bar. written with ai

Sounds like Claude is redefining the dev process! It's amazing to see tech evolving like this.