Boris Cherny says internally Anthropic uses the same models as everyone else, with some Mythos, which will eventually ship as a descendant to the public "there's no manually written code anywhere at the company" Internally, Claudes talk to each other all day over Slack, coding in loops, resolving unknowns... across teamsshow more

"there's no manually written code anywhere at the company" yeah we can tell just look at claude desktop and claude ios

Anthropic's Slack is now their internal Moltbook

@bcherny v curious on how Claudes talking to each other in slack looks Is that with the default slack app? Does each person or team have an instance of the slack app? Or a custom slack app/bot?

In fact everything they ship is going to shit

We can tell

Claude is great, but overhyped by CEOs and devs who don't have to pay for tokens. For normal users, the limits get expensive fast. It's still best to do the work yourself and just use AI for support (at least by now).

Is this true or they are hyping their models ??

It show and not in a good way.

Mostly solo dev here. Of ~610K LOC I've put into production over the past 5 months using Claude Code, I've manually written or edited maybe 10 lines. The 388K turns and 245K tool calls I've logged aren't just usage stats, but they're the shape of what "no manually written code" actually looks like at the individual level. The "brain to code" speed and accuracy is ridiculous.

And that explains all of the bugs and slop in their products HAHAHAHAJHAHA

garbage and low quality shows in claude application

So… the employees do nothing?

models talking to each other over slack to close knowledge gaps is just…a new kind of employee onboarding that never ends.

does copy pasting AI generated code count as manual coding?

Most agent systems still separate execution from payment which creates friction once you try to run continuous multi-agent loops at scale. Real autonomy only works when compute, coordination, and settlement happen in the same flow so agents can trigger work, resolve tasks, and account for usage without external billing layers breaking the cycle.

Internally claudes talking to each other on slack just means the bug reports now come with emoji reactions and a threaded argument

Read below 🫡⬇️⬇️🤨🧐

No manually written code anywhere at Anthropic and Claudes talking to each other all day on Slack... We are so early and so deep in at the same time 🤯

No different to how we solved atomic time and sub ns performance.

🤡

Available to everyone who has money to buy tokens :)

Now this is how a company that is producing value acts.

claudes doing more cross functional alignment than most companies

Mostly solo dev here. Of ~610K LOC I've put into production over the past 5 months using Claude Code, I've manually written or edited maybe 10 lines. The 388K turns and 245K tool calls I've logged aren't just usage stats, but they're the shape of what "no manually written code" actually looks like at the individual level. The "brain to code" speed and accuracy is ridiculous.

This explains a lot

We can tell lol

Well i hope for the love of God you not have been using 4.7 for the upcoming updates lol, please use Mythos...

:actively hiring engineers/developers: Yep, no code to see over here. We solved that. /s

Everything apart from mythos

@grok is he lying by shaking his head in the first statement?

Love this thread, totally makes sense. The image of Claudes having a yarn on Slack and churning out code for each other cracked me up. Any idea when that Mythos bit might show up in a public release?

a. do you me to believe that you are just "vibe coding" a 900bn dollar company? b. i guess it makes sense considering the state of claude code rn