Boris Cherny, the engineer who built Claude Code at Anthropic, said the quiet part out loud at Sequoia's AI Ascent: coding is solved. He hasn't written a line of code by hand in 2026. He runs a few dozen PRs a day. He hit 150 PRs in a single day last week — not as a job, as a stress test. Most of his work happens from his phone, with hundreds of agents running in the background and a few thousand more grinding overnight.
If that sounds like a flex, it isn't. It's a description of where the floor is now, not the ceiling.
The number that should rattle every product team is this: Claude Code was built six months before product-market fit, deliberately. Anthropic shipped it knowing it would be barely useful at launch because they were building for the next model, not the current one. That bet — the "product overhang" bet — paid off when Opus 4 dropped in May. Then 4.5. Then 4.6. Then 4.7. Each release inflected the curve. The product didn't catch up to the model; the model caught up to the product.
That's the playbook now. Build six months out, not for what the model can do today.
The real shift isn't speed. It's loops.
Cherny's most interesting comment wasn't about how fast he codes. It was about what he does instead of coding. He has dozens of agents running on cron — scheduled jobs that babysit pull requests, auto-rebase, fix flaky CI, cluster his Twitter feedback every thirty minutes. Anthropic just shipped "routines" so these can run server-side after you close your laptop. Inside the company, his agents talk to other people's agents on Slack to resolve unknowns without humans in the chain.
This isn't faster typing. It's a different topology. The unit of work stops being a task and becomes a standing process that runs forever and improves itself. Most engineers and most founders are still thinking in tasks. The few who are thinking in loops are pulling away.
The SaaS apocalypse take is more interesting than the meme version.
The lazy take is that AI eats SaaS. The Cherny version is sharper. Borrowing from Hamilton Helmer's seven powers, he argues two of them get weaker and the rest don't. Switching costs go down because models port your data and integrations between tools cheaply. Process power — workflow software whose moat was "we encode the messy steps" — collapses, because the model itself is now extremely good at figuring out and hill-climbing any process you describe.
What survives: network effects, scale economies, cornered resources, brand. The structural moats. The earned moats. The moats that don't come from making your customer's life painful to leave.
The corollary is the part founders should screenshot. Cherny thinks the next ten years will produce roughly 10x more disruptive startups than the last ten, not because building is easier but because incumbents are stuck. They have to retrain entire orgs, fight internal resistance, and migrate workflows that thousands of people depend on. A two-person team starting fresh today can build AI-native from line one and aim straight at companies worth a hundred billion dollars. The asymmetry is enormous, and it's only open for a window.
The deepest claim in the talk is the printing-press parallel.
In the 1400s, about 10% of Europe could read and write. Within fifty years of Gutenberg, more literature was published than in the previous thousand combined, and the cost of a book fell roughly 100x. Today, global literacy sits near 70%. Reading and writing went from a paid profession to a baseline life skill — though there are still professional writers, and they're very good.
Cherny thinks software is on exactly that arc, faster. Not "no-code for non-engineers." Software creation as a default literacy. The best person to build accounting software, in his framing, is a really good accountant — not an engineer — because the domain is the hard part and the coding is the easy part. Same for legal. Same for logistics. Same for whatever vertical you happen to know cold and the engineers don't.
If that's right, the most valuable people in every industry over the next decade are not the AI specialists. They are the deep domain operators who pick up coding the way previous generations picked up Excel.
What this means if you're building.
Three things. One: stop pricing your roadmap by what the model can do today. Build for the model six months out, accept the dip, and be ready when the inflection lands. Two: stop thinking in tasks. Start thinking in loops, routines, agents-talking-to-agents. The product surface is changing under you. Three: if your moat is switching costs or process complexity, it's already evaporating. If you don't have a real network effect or a structural advantage, you have months, not years, to build one.
The window is open. It will not stay open forever. The people who already write zero lines of code are not waiting for permission.
No comments:
Post a Comment