Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next

ELI5/TLDR

Boris Cherny is the engineer who built Claude Code inside Anthropic. He says coding, for him at least, is “solved” — the model writes 100% of his code, he ships dozens of PRs a day from his phone, and he keeps a few hundred agents running in loops at any given time. The rest of the world is behind, but the gap is shrinking. The bottleneck moved from typing code to figuring out the domain, the process, and the organization around it.

The Full Story

The accidental origin

Cherny joined Anthropic Labs in late 2024 — a small incubator that quietly produced Claude Code, MCP, and the desktop app, then disbanded. The thesis was a “product overhang”: the model could already do far more than the existing tooling let it. In late 2024, the state of the art was tab-completion — Sonnet 3.5 could finish a line. Cherny’s bet was that the next model could write the whole thing, agentically, no IDE.

So, we don’t have to do type ahead anymore, we can just have the agent write all of the code.

It didn’t work. For six months it was barely usable — he got maybe 10% of his own code from it. The first public release also flopped, in the sense that it had users but no exponential. Inflection arrived with Opus 4 in May 2025, and every model release since (4.5, 4.6, 4.7) bent the curve up again. The whole project was pre-PMF on purpose, built for a model that didn’t yet exist.

What “coding is solved” actually means

When Cherny says coding is solved, he means it for himself, in his codebase, with his harness — Claude Code itself is “just TypeScript and React,” chosen back when language choice mattered because the model needed to be on-distribution. By October–November 2025, the model wrote 100% of his code. He now ships “a few dozen PRs every day,” with a personal record of 150 in one day, mostly from his phone.

for me it’s just solved. Um but this is not the case everywhere. There’s very big complicated code bases. There’s kind of weird languages the model’s not good at yet… Usually the answer is just wait for the next model.

The room mostly raised hands at “somewhere in between” — neither all-by-hand nor all-agent. The honest read: Cherny’s “solved” is a leading indicator from a clean codebase in a popular language, not a description of where most teams sit today.

The personal setup: loops, not chats

His current workflow is the most useful concrete artifact in the talk. He runs five to ten Claude sessions, each spawning a few hundred agents in parallel during the day, a few thousand overnight. The thing he’s leaning on hardest is /loop — a tiny primitive where Claude schedules itself via cron to run on an interval.

All it is is you have Claude use cron to schedule a job for some point in the future, and it’s a repeat job.

His running loops, as examples:

One babysits his PRs — fixes CI, auto-rebases.
One keeps CI healthy — chases flaky tests.
One scrapes Twitter feedback every 30 minutes and clusters it.

Anthropic also just shipped “routines,” which is the same idea but server-side, so loops keep going after the laptop closes. His pitch: “loops are the future at this point.”

What stays human

Less than you’d expect, on his team. Cherny says everyone on the Claude Code team codes — engineering manager, PM, designers, data scientist, finance, user researcher. Specialists by title, generalists in practice.

The bigger shift he’s predicting is cross-disciplinary generalists — engineers who are also good at design, or product engineers who are also good at data science. The “T-shape” gets fatter on top.

The hard part stops being syntax and starts being domain.

The best person to write accounting software, I think maybe even today, is not an engineer, it’s a really good accountant because they know the domain really well and coding is the easy part.

The SaaS-apocalypse question

Asked whether cheap code kills SaaS, Cherny reaches for Hamilton Helmer’s 7 Powers framework. Two of the seven get weaker:

Switching costs — the model can port from one tool to another, so lock-in erodes.
Process power — Claude is now good enough at figuring out workflow and process that proprietary process know-how is less defensible. He singles out 4.7 as the first model that can “hill climb anything” — give it a target and it iterates until done.

The other powers — network effects, scale economies, cornered resources — still hold. They were never about code velocity.

The second half of his answer is the more interesting bet: a 10x increase in disruptive startups over the next decade, because incumbents have to retrain humans and rewire org charts, while a fresh team can build AI-native from day one.

The Anthropic-vs-everyone gap

Cherny is careful here. On the model, there’s no real gap — internal Anthropic uses the same models customers do, mostly Opus 4.7 with a bit of an unreleased model called “mythos.” The dog-fooding is the point.

The real gap is organizational. Inside Anthropic:

We have no more manually written code anywhere at the company. All of the SQL is written by uh by models. Everything is just built by the models.

Their Claudes talk to each other — his loops Slack-message his colleague’s loops to resolve unknowns. The frontier isn’t the tooling, it’s the willingness to rewire process around it. Startups have the easier time of it because they have nothing to unwind.

The printing press analogy

Asked whether building software becomes a普通-person skill, Cherny reaches all the way back to 1400s Europe. Pre-press, ~10% of Europeans were literate, often working as scribes for illiterate kings. In the 50 years after Gutenberg, more was published than in the previous thousand. Cost of a book fell ~100x. Two centuries later, literacy hit ~70% globally.

His claim: software is heading the same way, but faster than 50 years. There will still be professional engineers (just as there are professional writers), but writing software stops being a credential-gated skill.

Local vs. cloud, MCP vs. computer use

A questioner asked whether local models become the default once open-weights catch up. Cherny’s answer is essentially “shrug — the model decides.”

I don’t think these will be decisions that we are making as engineers anymore.

For knowledge-work tools that aren’t local-first, his answer is MCP for things that have an API, computer use for things that don’t. Anthropic is “pretty far ahead” on computer use; 4.7 made it slow-but-actually-usable. To the model, MCP and APIs and screen-clicking are all just tokens.

Key Takeaways

Build for the next model, not this one. Claude Code was deliberately pre-PMF for six months. The product was waiting for Opus 4.
The harness matters less as the model improves. Today’s safety scaffolding (prompt-injection guards, command verification, permission modes, human-in-the-loop) becomes vestigial once the model is aligned enough to “just do the right thing.”
Loops > chat. A cron-scheduled, self-restarting agent doing one job on a clock is the primitive that keeps coming up. Not multi-agent orchestration — just a recurring task.
150 PRs in a day. One developer, one phone. Useful as a ceiling on what’s already possible, even if most days are nowhere near that.
The 7 Powers update. Switching costs and process power weaken under AI. Network effects, scale economies, cornered resources do not. Pick moats accordingly.
10x more disruptive startups in the next decade. Incumbents drag because of org friction, not technology gaps.
Domain expertise becomes the bottleneck. A great accountant who can prompt is now the right person to build accounting software.
MCP for everything with an API. Computer use for everything without one.

Claude’s Take

Sequoia interviews come with a built-in vendor halo — Boris is on stage to make Anthropic look like the future, and most of the audience is already paying customers. So filter accordingly. The “coding is solved” framing is a marketing line that he himself walks back inside the same answer (“usually the answer is just wait for the next model”). And the 150-PRs-in-a-day flex is interesting as a stunt, less interesting as a workflow — most of those PRs are presumably small CI fixes and rebases, not the kind of design-heavy work that’s the actual hard part of engineering.

What’s worth keeping:

The loops idea is genuinely useful and not vendor-specific. A scheduled agent that does one boring job on a clock — keep CI green, cluster Twitter feedback, babysit PRs — is a much better unit than the more glamorous “spawn 10 sub-agents in a tree” pattern people keep showing off. Worth experimenting with on your own setup.

The 7 Powers re-read is the sharpest part of the talk. “Switching costs and process power get weaker, the other five don’t” is a clean, falsifiable claim that you can actually use to evaluate moats in any software business right now.

The org-gap-vs-model-gap distinction is also load-bearing. The interesting question for a startup isn’t “do you have access to the model” (you do) but “have you rewired your processes around it” (almost no one has). That part rhymes with how most companies failed to capture cloud-native or mobile-native value — the tools were available, the org-redesign wasn’t.

Score: 7/10. Better than a typical conference talk, light on hard data, heavy on a single practitioner’s view from inside the company that benefits most if his predictions come true. Worth the 25 minutes for the loops setup, the 7 Powers update, and the printing-press framing.