Claude Code creator, Boris Cherny: "Going from agents to loops is as big a jump as going from code to agents." 40 minutes of the head of Claude Code showing exactly how he works now. Loops running his code reviews, his maintenance, his fixes, on their own. Watch it,... then read the full breakdown of loops below.show more

Hilarious. 126 year old best practice used by scientists is suddenly all the rage to software engineers.

Another step towards autonomy.

agents were the beta, loops are the mainnet

@AnatoliKopadze loops are like having an autopilot for coding. more time to focus on actual problem solving now!

Boris Cherny legend for us i'm so enjoying to listen him everyday

40 min of actual workflow not theory not a pitch bro just showed his whole setup and left no excuses 🔥

when boris heard there were 107 questions he was definitely thinking about how he'd just use a loop to answer them all

ngl loops feel like coding found religion

Curious, isn’t that what all the coding harnesses currently already do for any prompt? They already make the changes, run the tests to verify the changes and make corrections. How is the loop different other than it is scheduled? Or if it is trigger based to start?

Right, the inner build → test → fix is the same. A loop just adds three things: it starts on its own (schedule or trigger), it keeps state across runs with fresh context each pass, and it only stops when a real check passes. The agent run is the body. The loop makes it run by itself until the goal is actually met, no human in the seat.

yeah, I ended up making a tiny “pre-flight checklist” for loops because I got tired of debugging vague goals after the fact

the agents-to-loops jump is real, but it raises the stakes on the stop condition. a loop that's slightly wrong just runs the mistake faster and for longer. is the hard part designing the loop, or knowing exactly when to break it?

That sounds groundbreaking! Excited to see how loops transform the whole process. Can't wait to dive into the details.

loops is just agents that got tired of waiting for permission introvert character arc honestly

Enjoyed this Q&A My takeaway: as AI gets us closer to generating most of the code, the real differentiator shifts upstream to the quality of the idea funnel — choosing the right problems, questions & oppty .That’s why the point on Fable’s limitation in idea generation matters

Agent loops mark an equally transformative leap as AI coding agents; they enable fully unattended, self-correcting development pipelines and redefine engineers’ core workflow.

오 따끈한 영상

The loop abstraction is powerful precisely because it turns the model from a one shot code generator into a persistent process owner. Most current agent frameworks still treat each run as a stateless transaction.

Yeah Karpathy and Steinberger also preaching the same!

The thing that makes the loop safe is the exit condition. Have it run tests after each fix and bail if green-to-red, otherwise an unattended loop happily commits its own regressions.

ngl loops feel like coding found religion

loops are definitely the next level for automation

The insight is that you don't need a smarter agent, you need a dumber agent with a reliable retry loop. State persistence + fresh context each pass + a quality gate is what separates production from demo. The agent run is commodity. The loop is the product.

the pace of coding and AI workflows keeps speeding up every day

Makes me think we're still early on agents

Loops feel like the right abstraction until a bug causes one to drift quietly for 48 hours before anyone notices. What does your monitoring look like when a loop goes sideways in prod?

Saving this, the agents to loops framing is the kind of thing that clicks once you see it

"Just use best models and focus on better output" - one very powerfull idea from Boris. ROI is so much better that time wasted on rrying cheaper models might not be best ROI on you time spent.

every level shift here rewires how you actually build. loops sounds like another one.

loops eating agents for breakfast rn we just kept adding abstraction layers til the thing started doing its own laundry

Loops turn the agent from a one-shot tool into a persistent background process. The hard part is no longer the prompt. It is defining the termination conditions and the audit surface.

in my day we just wrote the code ourselves...

the part that gets me is its not a demo its just how he actually works now when the guy who built the thing changes his own workflow u kinda have to pay attn ngl

those who ignore this will never get anywhere must seen interview indeed good job

Unlimited tokens and "I use Fable for everything". I don't think we can all relate to Boris. We need someone to show us how to work with a token quota and economically maximize all models for our projects.