Claude Code creator: “100% of my code has been written by Claude since Opus 4.5. most of my coding routine is automated by loops & dynamic workflows" In this 40-minute video, Boris shares how to build a self-improving agentic setup from scratch. Loops + dynamic workflows + auto mode... + routines - that’s the secret. Worth more than a $500 vibe-coding course. Watch today, then read below to learn how to build the same workflow.show more

100% written by the model is a real milestone, but written and owned are different things. what decides if this scales isn’t who typed the code, it’s who can fix it at 2am when a loop quietly does the wrong thing. self-improving is great right up until it improves in a direction nobody is watching.

learning this from the person who built it actually matters

This watch is the best breakdown on his coding setup

The role of a developer is slowly becoming orchestration.

thanks for sharing this I saw the Codez

You are welcome Gip ! Super useful video for me

Boris Cherny goat, love to reading and listen something from him

thanks for sharing!

You are welcome bro. Hope it’s useful for you

auto mode sounds dangerous useful

Yeah. Auto is just next level. But loop engineering is future!

"100% since opus 4.5" at least names a starting point. better hook

Loops are great for code. For the rest of the business the part that compounds isn't the loop, it's the memory it reads and writes. I run marketing, decisions and clients as files the agent updates after every task. Boring, but it never forgets and never re-asks.

The future isn't writing better prompts. It's designing better workflows that let agents keep working without constant supervision. 🚀

this is the future of coding, honestly feels like cheating now

This sounds less like engineers being replaced and more like engineers becoming workflow designers. Where does the harness break down most often: ambiguous specs, visual QA, or cross-team handoffs?

This is impressive work — loops, dynamic workflows, and self-improving agents are a big step forward. But there’s a much deeper layer most people haven’t touched yet. The real leap isn’t just automating code with loops. It’s building recursive, multi-domain intelligence — multiple specialized systems running in parallel, constantly cross-referencing each other across completely different fields, all feeding into a central core. We’re not just talking about better coding tools anymore. We’re talking about building something that actually starts to resemble how the human minds naturally operate. Most aren’t ready for that conversation yet.

bold claim but self-improving how exactly if the loop hits a bad pattern it just confidently automates itself into a corner 😐

Writing all the code with Claude is the headline. The real story is who did it: the person who knows exactly where the tool breaks. The loops aren't magic. They're failure modes he already mapped, turned into guardrails. Copy the loops without the map and you ship the confusion too.

That’s another way of him saying let’s burn tokens everyone and let a.i generate more a.i slops without human intervention and later down the road will be deleted because it generates enormous number of slops 🙂‍↕️

The sub agents part is the real unlock. One agent writes, one verifies, one explores. The model that wrote the code should never be the one grading it.

bro escaped the matrix and the matrix is just him typing 'make it better' in a loop 💀

最有说服力的是创造者自己把 100% 代码都交给 Claude + loops + dynamic workflows。 这说明真正的生产力革命，不是模型更聪明，而是把 coding 过程彻底变成可循环、可自优化的 agentic 系统。 未来写代码的人，本质上是在做 workflow engineering，而不是 prompt engineering。

That's impressive! Automating coding routines really takes things to the next level. Can't wait to check out that video and learn more!

been running this loop setup for a few months, the thing that wasn't obvious until we hit it: checkpoints where the agent has to flag uncertainty before it continues. without that you get confident-wrong output at scale and it's hard to catch after

Boris is everywhere lately 😂

w content

This is exactly where the industry is moving. Anyone still manually copy-pasting code snippets into a chat window is living in 2023. ​The magic isn't just asking an AI for code; it's building the autonomous loops that let it test, fail, and fix itself without you babysitting it

40 mins of actual signal for free while people pay $500 for fluff boris said let the loop cook and walked away 😭

i'm genuinely in doubt if that's a good thing or bad cuz i'm using claude code and sometimes desktop app and i'm tired of complaining about the bugs :) and at anthropic scale the downtime is just not what is expected

loops get the hype memory + verification get the results without disk state and a check step you're just automating manual prompting on a timer

this Cherny interview is packed with solid practical advice

Boris đã chia sẻ bí quyết xây dựng thói quen agentic.

The subagent harness makes sense for engineering throughput, but Canva’s hard part is still product taste. Where does human review sit in the loop when the output affects design quality?

Claude is amazing, and anthropic is helping shape the US policy around frontier AI models.

learn how to prompt < learn how to make automation loops

100% of code being written by Claude since Opus 4.5 is a pretty bold claim worth digging into the actual workflow behind 👀