Boris Cherny hasn't written code by hand in eight months. Everyone already knows Claude Code can write code. The real story is that Anthropic is learning how to run coding agents while you sleep.
Cherny built Claude Code. Now he mostly manages it. At Fortune Brainstorm Tech in Aspen on June 8, he told Fortune's Jeremy Kahn that he hadn't written a line of code by hand in roughly eight months, and that Claude Code itself is now fully written by Claude Code. That's a hard claim to ignore, even in an industry that overclaims for a living.
Fortune reported that Cherny had been managing a few hundred agents on the morning of the talk, and that on some days the count rises into the thousands or tens of thousands. This is the part you should pay attention to. The shift isn't from human coding to one chatbot writing a function. It is from a person prompting an assistant to a system where Claude instances prompt other Claude instances, review pull requests, scan for security problems, and wake up with work already done.
That sounds absurd until you look at the workflow. Cherny isn't typing one perfect prompt and waiting for a miracle. He is designing loops: recurring agent runs with a trigger, a task, a verification step, and a reason to stop. TechCrunch reported on June 22 that at Meta's @Scale conference, Cherny called loops as important a step as the move from source code to agents. He was right to put it that strongly.
Software work has always had dull repeatable edges. Pull requests stall. Tests fail. Review comments get missed. Architecture drifts because no one wants to spend Friday afternoon hunting duplicated abstractions. Cherny's point is that those jobs can now be handed to loops that keep coming back until the work is either done or clearly needs a human. That's different from autocomplete. It is operations.
The catch is verification
Don't mistake this for magic. Cherny's own examples work because the agents have checks around them. If a loop can run tests, inspect a real environment, compare screenshots, or prove that a pull request satisfies a goal, then you have a workflow. If it can't, you have an unattended guess running up a bill.
Fortune's June coverage also carried the more useful caveat: Anthropic says its code output has risen sharply, including an 8x increase by one company measure, but the company itself has warned that code volume is a rough proxy. More lines aren't the same as better software. Anyone who has reviewed a bloated pull request knows that already.
Still, the bottleneck has plainly moved. Cherny said Anthropic responded to automated code generation by automating review, using teams of Claude agents with different perspectives to look for bugs. That doesn't remove engineers from the process. It changes the job. You become the person who defines the finish line, checks the evidence, and decides when a machine-written change deserves to ship.
That's where most teams are not ready. If your test suite is weak, your issue descriptions are vague, and your review culture depends on one senior engineer remembering every hidden rule in the codebase, a thousand agents won't save you. They will multiply your confusion. Frankly, that is the immediate risk for engineering leaders buying into the agent story too quickly.
Kimi shows the other race
This is landing while the model race is getting stranger. The Associated Press reported today that Moonshot AI's Kimi K3 has surprised the U.S. tech industry with front-end coding results that rival top systems from Anthropic and OpenAI. Business Insider noted that the 2.8 trillion parameter open-weight model reached the top spot on Arena's frontend coding leaderboard while still trailing leading proprietary models on broader intelligence rankings.
Kimi K3 proves the point: better base models will keep arriving from more places. Moonshot, DeepSeek, OpenAI, Anthropic, you name it. The old question was which model writes the best answer to a prompt. Cherny is asking a colder question: which company can turn model output into reviewed, verified, repeatable work at scale?
That is why the Claude Code story matters. Anthropic isn't only trying to win a benchmark table. It is trying to own the system around the model: the terminal, the cloud routine, the review agent, the loop, the handoff back to a human. If that layer works, the advantage is not just smarter code generation. It is less idle time between intent and shipped work.
You don't have to believe the tens of thousands of agents number will become normal for every team. It probably won't. But you do have to take the direction seriously. The best engineers are already moving from writing code to managing work systems that write, test, and review code for them.
The uncomfortable part is that this rewards disciplined teams first. Clean tests, clear ownership, good review gates, and explicit acceptance criteria used to sound like process hygiene. In the agent era, they are the rails. Without them, overnight automation is just a faster way to create morning cleanup.
Also read: Meta Is Reportedly Negotiating to Rent Its AI Chips to Rival Anthropic • ASML Gives Nearly Every Worker 20,000 Euros in Stock to Keep Them Through 2030 • Kimi K3 Shows the Cheap Chinese AI Era Is Coming to an End
