Coding Is Solved. What Comes Next?

In Brief

One Tuesday last week, Boris Cherny submitted 150 pull requests in a single day. From his phone. He didn't write a line of code himself. It wasn't a goal, more an experiment to find the ceiling. On a normal day, it's a few dozen. The number isn't the interesting part. What is: what it tells us about what's actually solved, and what isn't.

From accidental experiment to tool everyone uses

Boris Cherny came to Claude Code almost by accident. In late 2024, he joined Anthropic Labs, a small innovation unit inside Anthropic. The team was tiny, and they built three things: Claude Code, MCP (Model Context Protocol, the standard for connecting AI to external tools), and the desktop app. Then the team disbanded.

The first six months, Claude Code didn't work well. Boris used it for maybe 10 percent of his code. After launch, there was no breakthrough: lots of users, but no exponential growth. Then Opus 4 arrived in May 2025. That's when things took off. Growth has inflected with every model release since: Opus 4, 4.5, 4.6, now 4.7.

The team has since reassembled for round two, led by Mike Krieger, Anthropic's chief product officer and one of the co-founders of Instagram.

What "solved" actually means

When Boris says coding is solved, he doesn't mean all coding challenges are gone. He means that for the code he writes (TypeScript and React, well represented in the model's training data) it's no longer a question. The model writes it. 100 percent.

For large legacy codebases in less common languages, challenges remain. "Usually the answer is just wait for the next model," he says.

The Claude Code codebase was leaked and isn't particularly mysterious: TypeScript and React, fairly straightforward. Choosing those languages was a deliberate call. When the project started, the model was best trained on them, and they wanted to minimize friction. Today the model can handle most languages, but in 2024 it mattered that the codebase was on distribution, meaning in the center of gravity of what the model had seen during training.

What a workday looks like

Boris now works primarily from his phone. Inside the Claude app there's a tab for code sessions. He keeps 5-10 sessions running at once, each with a handful of agents. Together that's a few hundred agents during the day, and a few thousand at night when deeper work happens without anyone waiting.

The thing he's found most valuable is /loop. It's a Claude Code command that uses a schedule (cron) to run a task repeatedly: every minute, every five minutes, every day, however you want. Boris now has dozens of these loops running:

• One babysits pull requests and fixes build failures automatically
• Another keeps the test environment healthy and fixes unstable tests
• A third pulls feedback from Twitter and clusters it for him every 30 minutes

"I sort of feel like loops are the future at this point," he says. Anthropic has also launched Routines, the server-side version of /loop that keeps running even when you close your laptop.

Internally at Anthropic, Claudes write code in loops and communicate over Slack with other users' Claudes also running in loops. Nobody writes database queries by hand anymore. Everything is built by the models.

The new problem isn't the code

At Anthropic, everyone on the Claude Code team now codes. The engineering manager. The product manager. The designers. The data scientist. The finance person. Everyone.

What Boris predicts is that we'll see far more cross-disciplinary generalists. Not generalists who know a bit of iOS and a bit of web, but people who are strong across product, design, and development at the same time. Code is no longer the barrier.

"The best person to write accounting software is not an engineer, it's a really good accountant," he says. They know the domain. Coding is now the easy part.

SaaS and the power shift ahead

Boris drew on Hamilton Helmer's framework "7 Powers" (seven types of competitive advantage in business) to explain what's happening to the software industry. Two of those powers weaken with AI:

Switching costs weaken because AI makes it easier to move data and code between systems. Process power weakens because the model is now good enough to understand and replicate most internal workflows. With 4.7, it can hill-climb toward any target given enough iterations. Boris calls it the first model to do that.

Network effects, economies of scale, and control of scarce resources are largely unaffected by AI. They remain as powerful as before.

His second prediction: the number of startups disrupting established industries will grow tenfold over the next decade. A small startup can now build something as valuable as a large company. The large company has to retrain employees, change processes, overcome internal resistance. The startup starts from scratch and builds AI-natively from day one.

The printing press and what comes next

Boris reads a lot of tech history. The historical parallel he reaches for is the printing press, invented in Europe in the 1400s.

Before it, roughly 10 percent of the European population was literate. They worked for kings and lords who couldn't read themselves. In the 50 years after the first printing press, more literature was published in Europe than in the preceding thousand years. The price of books fell roughly a hundredfold. And over the next few centuries, global literacy rose from 10 to 70 percent.

Reading and writing became something everyone can do. Professional writers still exist. But you don't need a degree in writing to write.

Boris believes programming is heading the same way. Just much faster than 200 years. Coding will become a skill like sending a text message. Not something only people in tech do. And in an industry where coding is the easy part, what you know about accounting, medicine, logistics, or law is what sets you apart.

---

Glossary

---

Sources and resources