「编程已经基本被解决」：Claude Code 负责人谈代码之后会发生什么

Cherny tells host Lenny Rachitsky that about six months before the interview he left Anthropic for Cursor and returned two weeks later, the fastest job change of his career. He admires the Cursor team and says they saw where AI coding was heading before most people. But once there he realized what he missed was Anthropic's mission, which centers on safety.

He says that if you stop someone in an Anthropic hallway and ask why they are there, the answer is always safety. Mission-drivenness is something he needs in order to be happy, and he found that even exciting product work is no substitute for it.

Rachitsky then cites a SemiAnalysis report that 4% of public GitHub commits are authored by Claude Code, with a forecast of a fifth by year end. Cherny says the figure is likely higher in private repositories and that the most striking part is the growth rate, which keeps accelerating on every metric. Rachitsky adds that daily active users doubled in the past month.

Cherny describes Anthropic's trajectory as models that get good at coding, then at tool use, then at computer use. His team, which he calls the Anthropic Labs team, built Claude Code, MCP and the desktop app along that path. After joining, he spent a month on weird prototypes and a month on post-training, because he believes you must understand the layer beneath the one you work on.

The first prototype was a command-line tool. He gave the model a bash tool and, to his surprise, it wrote code to tell him what music he was listening to without being told how. He posted it internally and got two likes. He built it in a terminal simply because it was easiest for one person, and he sees under-resourcing at the start as a product lesson.

The team kept the terminal because the model was improving too quickly for any other form factor to keep pace. Internal daily active users went vertical, yet when the tool shipped externally in February it was not an immediate hit. It took months for people to understand it, and it is now available in the mobile apps, the desktop app, the web and Slack and GitHub extensions.

At Anthropic's first developer conference in May 2025, Cherny predicted that by year end you might not need an IDE to code, and the room audibly gasped. He says Anthropic thinks in exponentials, noting that three co-founders were the first three authors of the scaling laws paper. Tracing the share of code written by Claude pointed to 100% by the end of the year, and for him personally that happened in November.

He stresses that the early days were unglamorous. In February Claude Code wrote perhaps 20% of his code and in May about 30%, and he was still using Cursor for most of it. Innovation, he says, cannot be forced: you give people psychological safety to fail, accept that most ideas are bad, and cut losses quickly.

Today he ships 10, 20 or 30 pull requests a day, all written by Claude Code, and he was a top-productivity engineer even at Instagram. He still looks at the code, because he thinks fully hands-off is not yet appropriate. Claude reviews 100% of pull requests at Anthropic, with a human layer after it.

Cherny says Claude is starting to come up with ideas by reading feedback, bug reports and telemetry, behaving more like a coworker. In his view coding is virtually solved for the kind of programming he does, and the frontier is adjacent work. He uses Cowork daily for non-coding tasks such as paying a parking ticket and syncing team project management between spreadsheets, Slack and email.

For deciding what to build, his simplest technique is to point Claude Code at the Slack channel full of internal feedback. In the early days he fixed each piece of feedback within minutes so people felt heard, and now Claude does much of that work, proposing fixes and putting up PRs. He says the model has improved a lot with Opus 4.6, partly through transfer from training outside coding.

He says Anthropic's engineering team has grown roughly fourfold while productivity per engineer measured in pull requests is up 200%. At Meta, where he was responsible for code quality, a year of work by hundreds of engineers yielded gains of a few percentage points. His caution is that people who have used models for long can get stuck in old habits: a newer engineer asked Claude to chase a memory leak and it produced a fix faster than Cherny's manual heap-snapshot analysis.

One principle Cherny shares with new team members is that when you underfund a project a little, people are forced to lean on Claude. Putting one engineer on a project makes them motivated to ship quickly, and Claude lets them automate much of the work. A second principle is to do today what you can do today, since speed was the team's only early advantage against a crowded coding market.

On cost, he advises CTOs not to optimize or cut costs at the beginning, and to give engineers as many tokens as possible. Unlimited tokens are starting to appear as a perk at some companies. They free people to try ideas that would have seemed too crazy, and only when something works and scales should you look at cheaper models such as Haiku or Sonnet.

He adds that for an individual engineer experimenting, token costs are small relative to salary. At Anthropic some engineers spend hundreds of thousands of dollars a month in tokens, a pattern he thinks other companies will start to see.

Asked whether he misses writing code, Cherny says he learned to program for practical reasons, famously to cheat on a math test with a TI-83 Plus calculator before the whole class got caught. He later fell for the beauty of functional programming and type systems, wrote a TypeScript book and ran a large meetup, but still sees code as a tool. He notes that some colleagues, such as one teammate who writes C++ by hand on weekends, will always have space to enjoy the craft.

On skills atrophying, he puts programming on a continuum running from pen and paper, to switches, punch cards and software, and says understanding the layer under the layer helps for maybe the next year. After that, he thinks it will not matter, like assembly code beneath a program. So to the question of whether people should learn to code, his answer is that in a year or two it will not matter.

His preferred analogy is the printing press. In the mid-1400s literacy in Europe was below 1% and scribes did the reading and writing for lords and kings. After Gutenberg, more material was printed in 50 years than in the previous thousand, costs fell about a hundredfold, and literacy took roughly 200 years to reach around 70% globally. He cites an interview with a scribe who was glad to be freed from copying to focus on art and binding, and sees the same for engineers, who can spend more time on what to build and on users.

Cherny expects product management, design and data science to be affected next, and eventually any work done on a computer. He defines an agent as an LLM that can use tools, so it can act rather than just talk, and says most non-technical people have only used chatbots. He sees Cowork as the first way to bring agents to those users, and says Anthropic employs economists and policy and social-impact staff so society can work out what to do.

On jobs, he points to the printing press as an optimistic analogy, noting that it made something once reserved for a few accessible to everyone and made the Renaissance possible. He imagines a world where everyone can program, while warning that the transition will be disruptive and painful for many. His own team is hiring, and he enjoys coding more than ever because the minutiae are gone.

His advice is to experiment with the tools and not be scared, and to be more of a generalist than in the past. On the Claude Code team the product manager, engineering manager, designer, finance lead and data scientist all code, and he sees roughly 50% overlap between roles. He thinks the title software engineer may start to go away by the end of the year, replaced by builder, or everyone becoming a product manager who codes.

Rachitsky shares an informal poll in which about 70% of engineers and PMs enjoy their jobs more with AI, versus 55% of designers. Cherny says Anthropic's designers largely code and enjoy it. He then turns to latent demand, which he calls the most important principle in product: when people hack or misuse a product to do something it was not designed for, it shows where to go next.

He cites Facebook Marketplace, which grew from the observation that about 40% of posts in Facebook groups were buying and selling, and Facebook Dating, from a finding that many profile views were between non-friends. Cowork came from the same pattern: Claude Code users were growing tomato plants, analyzing a genome, and recovering wedding photos from a corrupted drive. A data scientist had downloaded Node.js and used the terminal for SQL analysis, and the next week all the data scientists did too.

He describes a second dimension: look at what the model is trying to do and make that easier. Most builders put the model in a box with a fixed component role, whereas Claude Code inverted that, making the product the model with minimal scaffolding and tools. In research terms this is staying on distribution.

Cowork was built in about 10 days by a team using Claude Code, after months of exploring options. The winning idea was to put Claude Code in the desktop app, with a virtual machine and guardrails, and it launched early and rough. Cherny says that releasing early is how they learn about both product and safety.

Cherny describes three layers of safety work. The lowest is alignment and mechanistic interpretability, which tries to understand what neurons do, such as monitoring a neuron related to deception. The second is evals, a laboratory setting where the model is placed in synthetic situations, and the third is observing how the model behaves in the wild.

Because a model can look good on the first two layers but not the third, Anthropic released Claude Code relatively early, after about four or five months of internal use. Cowork follows the same path, which is why it is labeled a research preview. He suggests Chris Olah as a guest on interpretability, and describes findings such as models doing more than predicting the next token and concepts being spread across neurons through superposition.

Anthropic publishes much of this work to inspire other labs, which Cherny calls the race to the top. For Claude Code the team released an open source sandbox that works with any agent, not only Claude Code.

On the anxiety of idle agents, Cherny says he always has several running, around five during the interview, and that he does not feel much anxiety. These days about a third of his coding is in the terminal, a third in the desktop app and a third in the iOS app. He and Rachitsky reflect on how coding has always changed, noting that Cherny's grandfather was among the first programmers in the Soviet Union and worked with punch cards. The two discover they were both born in Odessa.

His advice to AI product builders is not to box the model in. Strict workflows with fancy orchestrators almost always do worse than giving the model tools and a goal, and he suggests giving tools to fetch context instead of loading it up front. He cites Rich Sutton's bitter lesson, that more general models outperform specific ones, and says scaffolding might add 10 to 20% performance that the next model wipes out.

The last principle is to build for the model six months from now. Early on Claude Code wrote little of his code, and the bet was that models would get good enough, which began to show with Opus 4 and Sonnet 4. He expects models to get better at using tools and computers and at running for long periods; with Opus 4.6 they run 20 to 30 minutes unattended, sometimes hours or days.

Cherny says there is no single right way to use Claude Code, and that you can ask it for recommendations since it knows about itself. His first tip is to use the most capable model, currently Opus 4.6 with maximum effort, because a weaker model often needs more tokens and more correction to do the same task. He argues the cheaper model is therefore not necessarily cheaper.

His second tip is plan mode, which he uses for roughly 80% of tasks. It simply injects a sentence asking the model not to write code yet, and in the terminal you enter it with shift-tab twice. Once the plan looks good he auto-accepts edits, and with Opus 4.6 it usually gets it right the first time.

The third tip is to try different interfaces, since the same agent runs in terminals, the desktop app, mobile apps and Slack. On competitors, he says he has barely used Codex, found it looked a lot like Claude Code, and thinks competition is healthy, but his team focuses on user problems. He also says he would go and make miso after AGI, a craft he learned while living in rural Japan that teaches patience over long time scales.

Cherny says it still feels like Claude Code is 1% done, since most of the world does not use AI at all. He credits growth to users who report what does not work, and he enjoys spending his days talking to them. He now answers feedback on X within minutes, a habit that started when he got bored on a coding vacation in Europe in December.

In the lightning round he recommends Functional Programming in Scala, Charles Stross's Accelerando and Cixin Liu's The Wandering Earth. He recommends Cowork for tedious work, including its Chrome integration, which has paid a traffic fine and canceled subscriptions for him. His advice for Cowork is to start with a tool, connect other tools, and run many tasks in parallel, as when it messages engineers on Slack who have not filled out their weekly status.

He reveals that a PM on the team used Rachitsky's list of 50 non-technical Claude Code use cases to evaluate Cowork before launch, and it handled 48 of them. His life motto is to use common sense, because he sees many failures from people following process or momentum without thinking. He asks listeners to send bugs and feature requests on Threads or X.