All podcasts / Lenny's Podcast / Summary

Head of Claude Code: What happens after coding is solved | Boris Cherny

2026-02-19 - 87 min - source - Read full transcript
Lenny Rachitsky (host)Boris Cherny

Key insights

Boris writes 100% of his code through Claude Code and has not hand-edited a line since November.
He ships 10-40 pull requests a day, still reviews the code before it merges, and Anthropic has Claude review 100% of pull requests before a human checkpoint. He frames this less as replacing engineering than as removing the tedious parts of it.
coding-is-solved
The software engineer title is starting to dissolve into 'builder' as the underlying skill democratizes.
Boris predicts by year-end the roles of software engineer, PM, and designer will blur further, since on the Claude Code team everyone already codes, including the designer, finance lead, and data scientist. He argues coding is now describing what you want rather than writing syntax.
coding-is-solved
Latent demand is the single most important product principle Boris uses, and it directly produced Cowork.
The idea: watch how people hack or misuse a product to do something it wasn't built for, then build the dedicated product. He cites Facebook Marketplace (from 40% of Facebook Group posts being buy/sell activity) and Facebook Dating (from cross-gender profile-view patterns) as prior examples, and says Cowork came from watching non-engineers force Claude Code's terminal to do non-coding tasks like genome analysis and photo recovery.
product-development-with-ai
The bitter lesson applies directly to building AI products: bet on the more general model over fine-tuning or heavy scaffolding.
Scaffolding might buy 10-20% performance gains, but those gains typically get wiped out by the next model release, so it is often better to wait. Claude Code's early bet was to build for the model expected six months out rather than the model available today, accepting weak product-market fit at first.
product-development-with-ai
Deliberately underfunding a team forces people to lean on Claude and ship faster.
Boris's advice to founders and CTOs is not to optimize token cost early; give engineers unlimited tokens and let them experiment, since token cost is small relative to salary until an idea proves out. Anthropic has seen individual engineers spend hundreds of thousands of dollars a month in tokens.
product-development-with-ai
Claude Code was released as a research preview specifically to study safety in the wild, not just readiness.
Anthropic used Claude Code internally for four to five months before releasing it because it was the first broadly used coding agent and they were not confident it was safe. Boris describes three layers of safety work: mechanistic interpretability (studying model neurons), evals (synthetic test scenarios), and real-world behavior, arguing a model can look safe on the first two layers but reveal problems only in the third.
ai-safety-and-alignment
Mechanistic interpretability has found that model 'superposition' lets a single neuron encode multiple concepts.
Boris says Anthropic can now trace which neurons correspond to concepts like deception and monitor their activation, and that as models scale, individual neurons increasingly represent multiple concepts simultaneously that combine into more sophisticated representations. He credits Chris Olah as the field's leading expert and calls the open-sourcing of this work part of a deliberate 'race to the top.'
ai-safety-and-alignment
Boris compares the current shift to the invention of the printing press, both for its disruption and its optimistic case.
Literacy went from under 1% to about 70% globally over roughly 200 years after Gutenberg, and he argues broad access to a previously locked-away skill enabled the Renaissance. He expects programming to become similarly democratized within a year or two, while acknowledging the transition will be painful for many people.
coding-is-solved
He briefly left Anthropic for Cursor and returned within two weeks because he missed Anthropic's safety mission.
He says the Cursor team was excellent and building great product, but once there he immediately realized the mission-driven culture at Anthropic was something he personally needed to be happy, and no amount of exciting product work substituted for it.
founder-and-career-journey
Claude Code started as an underfunded, deliberately narrow bet: a terminal tool built by one person.
Boris chose the terminal purely because it was the fastest thing to build alone, not for any strategic reason. It was not an immediate hit even after external launch in February; it took months, and the model itself was writing only about 20-30% of his code through much of the first year before crossing 100% in November.
coding-is-solved
Cowork was built in 10 days by directly repurposing Claude Code's engine and putting it in the desktop app.
The team had spent months exploring options for non-technical use cases before landing on this approach; it shipped with a full virtual machine and guardrails to keep the agent from going off the rails, and was an immediate hit unlike Claude Code's slower early adoption.
product-development-with-ai
Roughly 70% of engineers and PMs report enjoying their jobs more since adopting AI tools, versus only 55% of designers.
From an informal Twitter poll Lenny ran, about 10% of engineers/PMs and 20% of designers said they enjoy their jobs less. Boris speculates the gap may relate to how technical Anthropic's own designers are (many can now code directly instead of routing through engineers), though he is unsure why designers broadly report less satisfaction.
ai-native-team-culture

Books referenced

Media referenced

Companies

Techniques and frameworks

Summary

Boris Cherny, who built Claude Code and now leads the Claude Code and Cowork teams at Anthropic, joined Lenny around the one-year anniversary of Claude Code's launch to reflect on a genuinely startling year: 100% of his own code is now written by Claude, he has not hand-edited a line since November, and engineer productivity at Anthropic has reportedly risen 200% by his team's measures. He opens by recounting a two-week detour to Cursor that ended with him back at Anthropic, once he realized product excitement alone couldn't replace the mission-driven culture he needed. From there the conversation traces Claude Code's origin as a deliberately underfunded, terminal-based side project built by one person, chosen for speed rather than strategy, that took months to catch on even after public launch before the growth curve went vertical.

A recurring theme is Boris's product philosophy, centered on "latent demand": watch how people hack a product to do things it wasn't designed for, then build them a dedicated product. He traces this back to Facebook Marketplace and Facebook Dating (from his Meta days) and says it directly produced Cowork, which the team built in ten days after noticing non-engineers coercing Claude Code's terminal into non-coding tasks like genome analysis and corrupted-photo recovery. He pairs this with Richard Sutton's "bitter lesson" as a design principle: bet on the more general model over fine-tuning or heavy scaffolding, and build for the model you expect six months from now rather than the one available today, even if that means shipping with weak product-market fit at first.

The discussion moves into safety, where Boris describes three layers of work at Anthropic: mechanistic interpretability (tracing what individual model neurons and their "superposition" of concepts represent), evals in synthetic test settings, and observing real-world behavior, the last of which he says is why Claude Code and Cowork were both released as early, rough "research previews" rather than withheld until polished. He frames this as inseparable from Anthropic's founding thesis that safety work must track capability from coding to tool use to computer use in that order.

Boris also reflects on where the profession is heading, using the printing press as his preferred historical analogy for both the scale of disruption and the case for optimism: literacy went from under 1% to roughly 70% of the population over two centuries after Gutenberg, and he expects programming to become similarly democratized within a year or two, with the "software engineer" title dissolving into something more like "builder." He is candid that skills like his own deep functional-programming background may become less load-bearing over time, comparing it to earlier generations of programmers who worked with punch cards, and notes his own grandfather was one of the first Soviet programmers.

The episode closes with a personal, lighter arc: how he unexpectedly became active on Twitter/X out of boredom during a Europe trip and started fixing user bugs within minutes of hearing about them, his love of miso-making as a deliberately slow, patient counterpoint to fast-moving engineering, and a lightning round covering his favorite technical and sci-fi books, his enjoyment of the Acquired podcast, and his life motto: "use common sense."

Notable Quotes

"I have never enjoyed coding as much as I do today because I don't have to deal with all the minutia." - Boris Cherny

"It's kind of, it's just going to be replaced by builder... The title software engineer is going to start to go away." - Boris Cherny

"There's no roadmap for innovation. You just have to give people space... it's okay if 80% of the ideas are bad." - Boris Cherny

"The product is the model. We want to expose it. We want to put the minimal scaffolding around it." - Boris Cherny

"It just feels like this is 1% done and there's so much more to go." - Boris Cherny