Anthropic's Boris Cherny says Claude Code writes 100% of his code and he ships 10 to 30 pull requests daily.
Cherny says Claude Code alone makes $2bn in revenue and Anthropic $15bn.
Cherny predicts coding is virtually solved and by year-end everyone will be a product manager, everyone will code, and 'software engineer' becomes 'builder'.
Claude Code's creator says coding is nearly solved
Coding is close to solved inside Anthropic. Boris Cherny, who runs Claude Code, writes no code by hand, keeps several agents running and ships up to 30 pull requests a day. He argues that the roles adjacent to engineering change next, and predicts that by the end of the year everyone will be a product manager and everyone will code. Anthropic's answer is to build for the model six months out and to follow latent demand rather than invent use cases.
Takeaways
Coding is nearly solved. Boris Cherny predicts that by the end of the year everyone will be a product manager and everyone will code, with builder replacing software engineer. The roles next to engineering change first. He compares the shift to the printing press.
He writes no code by hand. Boris Cherny has not edited a line by hand since November and ships 10 to 30 pull requests a day, keeping about five agents running, sometimes from his phone. Productivity per engineer at Anthropic is up 200%.
Latent demand beats invented use cases. Boris Cherny calls latent demand the most important product principle, citing Facebook Marketplace, where 40% of group posts were buying and selling. Cowork came from people using Claude Code for tomatoes and wedding photos, and took ten days to build.
Agents now run unattended for minutes. A year ago Sonnet 3.5 could run for 15 or 30 seconds before going off the rails. Opus 4.6 now runs 10 to 30 minutes unattended on average, which is what lets Boris Cherny keep several agents going at once.
Claude Code alone makes $2bn. Lenny Rachitsky puts Claude Code's revenue at $2bn against Anthropic's $15bn. Claude reviews 100% of pull requests at Anthropic, with human review after, and the share of model-written commits is higher in private repos than public ones.
Safety work spans evals and interpretability. Anthropic open sourced a sandbox for running agents that works with any lab's agent, which Boris Cherny frames as a race to the top. Its safety work covers mechanistic interpretability, where he recommends Chris Olah, and evals.
How Claude Code was built
Boris Cherny spent a month hacking on prototypes, then a month doing post-training to learn the research side.
He built the terminal tool alone because no other form factor could keep up with a model improving that fast.
The internal announcement got two likes, and the public release in February was not an immediate hit.
Growth inflected with Opus 4 and again in November.
How Boris tells teams to build with AI
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Give engineers as many tokens as possible and do not box the model in.
Bet on the most general model, following The Bitter Lesson, and build for the model six months out.
Start about 80% of tasks in plan mode, which just tells the model not to write code yet.
Hire generalists. Anthropic screens every function for technical ability and its designers largely code.