Father of ClaudeCode: "In the era of 'everyone programming,' a company's real advantage lies in the generational gap of organizational processes."
Recently, Boris Cherny, creator of Anthropic's programming tool Claude Code, had an in-depth conversation with Lauren Reeder, partner at Sequoia Capital.
In front of a packed audience of tech investors, Boris stated that since October to November last year, all the code he writes has been 100% completed by models. By 2026, he hasn't personally written a single line of code.
For Boris himself, the issue of programming has been "solved". What enterprises truly need to compete on is the speed of organizational process reengineering.
His Workbench: Phone + Hundreds of Parallel Agents
Boris Cherny is a seasoned engineer, author of TypeScript programming manuals, and has written a vast amount of code in his career. However, he “hasn’t written a single line of code in 2026.”
Boris's main work device is his phone. He demonstrated his personal workflow onsite: opening the Claude mobile app, with a code tab on the left running five to ten sessions simultaneously, each session hanging a batch of Agents. “I usually write dozens of PRs every day; last week, I did 150 in one day—that was my record, just to see how far I could push myself.”
Boris said that currently there are hundreds of Agents running, and every night there are thousands working on deeper tasks.
He highlighted his most relied-upon workflow—Loop (loop scheduling):
This is the simplest yet most effective thing. You let Claude use cron to schedule a recurring task at a future time, which can run every minute, every five minutes, or daily.
He currently runs dozens of Loops: one continuously “watches” his PRs (auto-repair CI, auto-rebase); one maintains CI health (auto-fixes flaky tests); yet another fetches and clusters Twitter feedback every 30 minutes.
I now feel Loops are the future. If you haven’t tried it, I strongly recommend... Even if you close your laptop, it keeps running.
“The technical gap isn’t big; organizational gap is the real divide”
Someone asked: Compared to external developers, how many months ahead is Anthropic internally?
Boris's answer was unexpected. He said that at the model layer, Anthropic and outsiders are almost no different—“We use the same models. We place huge importance on dogfooding, since we’re building a platform, and developers must use the same tools as us.”
The real gap is elsewhere.
“I think where we’re ahead isn’t in technology, but in organizational structure and processes.” Boris said, “If you talk to anyone at Anthropic, you’ll find we use Claude for literally everything. My Claude runs loops to write code and communicates with colleagues’ Claudes via Slack to solve unknown problems. There’s no more handwritten code in the company. All SQL is generated by models, everything is built by models.”
This means: With the same tools, whoever transforms organizational processes first will have the real competitive advantage.
Universal Programming: Not Just an Engineer's Job
Boris Cherny explained that the Claude Code team itself is an experimental site for "everyone programs".
Our engineering manager, product manager, designer, data scientist, finance staff, user researcher—every member of the team writes code... Each has their own domain, but now everyone is programming at the same time.
He predicts, The future "polymath" will not only be cross-platform engineers (like doing iOS, web, and backend), but true interdisciplinary people: those who understand both product engineering and design, or combine data science with engineering.
For longer-term trends, he used a historical analogy. He reads two kinds of books: science fiction and tech history. "I think the most striking historical parallel now is the printing press in 15th-century Europe."
He described that before the printing press, only about 10% of Europeans were literate, employed by kings and lords. “Reading and writing was a specialized skill, not widespread.” After the printing press, in just 50 years, Europe published more texts than in the previous thousand years, and the cost of books dropped about 100-fold. After centuries, global literacy rose to about 70%.
His judgment is: “Software will become completely democratized. Everyone can do it, and it will happen much faster than the printing press.”
He predicts software development will become as ubiquitous as texting. He gave an example: “For accounting software, the most suitable person isn’t necessarily an engineer, but a highly skilled accountant—because domain expertise is the hard part, writing code has become the easy part.”
SaaS Landscape: Which Moats Are Disappearing
Asked whether AI will cause the “SaaS apocalypse”, Boris Cherny said it’s his “favorite question” and offered two views.
First, he believes AI will dilute two types of traditional moats:
- Switching Costs: Models can easily help users migrate between products;
- Process Power: Claude is increasingly good at replicating and optimizing complex business processes, especially Claude 4.7. “You give it a goal, tell it to iterate until completion, and it will do so. I think it’s the first model that can truly do this.”
Traditional moats like network effects, economies of scale, and scarce resources remain important and unchanged by AI.
Second, disruptive startups will emerge massively.
I think in the next 10 years, the number of disruptive startups will increase 10-fold. Now a small startup can build products equal in value to big companies and genuinely compete head-to-head—big companies must evolve business processes, retrain people to use new tech, which causes huge internal resistance. But you don’t have that problem. If you start from scratch, you can build everything natively with AI.
He concluded: “So I think now is the best time for entrepreneurship, the best time for startups, and opportunities for disruption are everywhere. We can achieve so much.”
The Future of Claude Code Itself
Asked what Claude Code will look like in a year, Boris Cherny gave a blunt answer:
I think Claude Code itself may only have 100 lines of code left in a year.
He explained that as models become increasingly autonomous, existing safety mechanisms—protection against prompt injection, static command validation, permission patterns, manual review loops—“will become less important because models will do the right thing themselves.”
He was also asked whether Claude Code’s success is due to the model or the product. He said, “Six months ago, maybe it was 50/50.”
He explained why the product layer is still important: “We pay huge attention to detail so you have a great experience using it all day.” But he also made clear, “As models get better, the ‘harness’ becomes less important.”
Full interview below:
Anthropic's Boris Cherny: Programming Is Solved—Where Do We Go Next?
Sequoia Capital · AI Ascent 2026Guest Introduction
Host (Lauren Reeder, Sequoia partner): Okay, I'm delighted to introduce our next guest. How many here use Claude Code?
(Hands raised)
How many have "Claude Code syndrome?"
(Laughter) Well, my team affectionately calls me as having Claude Code syndrome—maybe it’s true, maybe not.
We’re very happy to have Boris Cherny here. Boris is the creator of Claude Code and has witnessed and driven a comprehensive transformation in modern software development. Boris, thanks for taking the time to talk to us today. We know the future of software development rests, to some degree, on your shoulders. Our interviewer is Lauren Reeder from our team. Please welcome her.
(Applause)Chapter One: Claude Code User Survey
Lauren Reeder (Host): You stole my opener—I usually ask who uses Claude Code. Lots of hands, great.
Boris, thanks for coming, it's an honor. Everyone here is a founder or builder, and you’re changing the very act of “building.” I really want to explore your thoughts on the future of software development and what we should do with this “extra time.”
Before we dive in, a bit more background. Boris didn’t just create Claude Code; he’s really an engineer’s engineer—career spent writing loads of code, wrote textbooks like “Programming TypeScript.” When we last spoke, he said in 2026 he hadn’t written a single line himself—a huge shift.
One lesser-known fact: In middle school, I wrote a guide for TI-83 Plus calculator BASIC programming. I just looked it up—it’s still online, very embarrassing. Please don’t search for it.
(Laughter)
I’ll start with a few questions and leave lots of time for audience Q&A. Start thinking of your questions now.Chapter Two: Claude Code Origin Story
Lauren: Quick poll: among Claude Code users, who mainly uses the CLI? Ok, most are CLI. Who uses desktop? Okay. Who uses VS Code or JetBrains IDE? Not many. Any other way?
Boris Cherny (Anthropic, Claude Code Creator): I mainly use iOS now.
(Laughter)
Claude Code was, in some senses, an “accidental” creation. I joined an internal team, Anthropic Labs, at the end of 2024. The team completed its mission—built Claude Code, MCP, and desktop apps, just a few people, truly an innovation group. After the mission, the team disbanded, but now we’ve regrouped for a second round. Mike Krieger—you know him, Anthropic’s Chief Product Officer and Instagram co-founder—is leading this round.
I started building programming tools because we felt the so-called “product overhang”—many here know the term. We frequently used it internally; it basically means the model can do much more than any products currently leverage.Chapter Three: From Code Completion to Agents
Boris: In late 2024, the most advanced coding was “type-ahead completion”: you open your IDE and press Tab to autocomplete a line—Sonnet 3.5 first enabled this. But we sensed it could go further—the model was close to its next leap, no need for line-by-line completion, let agents write all the code directly.
So I tried it, but the first six months were rough—it was barely usable. I personally only used it for 10% of my code. When Claude Code launched, it didn’t explode; quite a few used it, but not the kind of exponential growth we see now.
The real turning point was Opus 4’s release in May—I remember it clearly. From then, exponential growth began, and each model release set another milestone—Opus 4, 4.5, 4.6, 4.7, all kept climbing.
We were essentially building for “the next model” before it launched, knowing there’d be no PMF for six months. But that was the plan. Anthropic has always been laser-focused, putting business, enterprise, safety, and programming at the center—this is our approach.Chapter Four: Is Programming Solved?
Lauren: The story is fascinating, especially the “accidental” part. You said publicly, “programming is solved.” Programming is one of Anthropic’s three big bets—can you explain what you mean? What problems remain unsolved? What second-order issues does this bring?
Boris: I’ll ask the audience: How many still handwrite 100% of their code? How many use Claude Code agents for 100%? How many are in between?
Good, about 50% solved.
For me, it’s 100% solved. Claude Code’s repo leaked; everyone knows it’s just TypeScript and React, nothing secret or particularly complex.
We chose TypeScript and React as they’re heavily represented in the model’s training data. At first, models weren’t so smart; language and framework choice mattered a lot. Now models can write any language, use any framework, even ones they’ve never seen. Back then, you wanted to stay in the mainstream.
So, we got to 100% model-written code very early—last October or November. Since then, it’s written all my code. I submit dozens of PRs daily. Last week, I submitted 150 PRs in one day—just to push the limits.
Of course, not all scenarios are solved. There are still huge complex codebases, rare languages models aren’t good at. But for my code, it’s solved. Generally, the answer is: wait for the next model release.Chapter Five: Boris's Personal Workflow
Lauren: Tell us about your personal setup—you showed me before, it was impressive.
Boris: I shared my workflow on Twitter six months ago; I was surprised people found it shocking—to me, it’s just normal coding.
(Laughter)
It’s evolved. Now, most of my work is done on my phone. Open Claude, the left has a “Code” tab with lots of live sessions. I usually run five to ten sessions, each with a batch of agents—hundreds running currently. Each night, thousands work on deeper tasks.
Managing agents has several approaches. One lets Claude call sub-agents for task allocation. But increasingly, I use “loops” (scheduled recurring tasks). It’s the coolest and simplest thing—let Claude use cron to schedule a task as often as you like.
I have dozens of loops running. For example: one monitors my PRs, auto-fixes CI, auto-rebases; one maintains CI health, auto-fixes flaky tests; one fetches Twitter feedback and auto-clusters every 30 minutes.
I really think loops are the future. If you haven’t tried them, highly recommended. We recently launched “Routines,” same principle but server-side—even if you close your laptop, tasks keep running.Chapter Six: The Team of the Future
Lauren: What is your vision for future teams? How should they operate—more agent collaboration or the current structure?
Boris: Tough to predict, but I’ll try. The big trend is: more “polymaths” in the future.
Now, “polymaths” basically means across engineering—someone does iOS, web, backend.
But I think we’ll see increasingly interdisciplinary polymaths—good at engineering, also at design; or understand both product and data science and engineering.
This is already happening in our team. Nearly everyone is an interdisciplinary polymath: engineering manager, product manager, designer, data scientist, finance, user research—all write code. Each has strengths, but all code.
I see some nodding—I suspect this isn’t surprising, as you may be experiencing the same changes.Chapter Seven: The Fate of SaaS
Lauren: Last round, then audience Q&A. We discussed programming changes, now about software products. With AI reducing the cost of coding by 10x or 100x, what happens to the value of software products? Will we see a “SaaS apocalypse?”
Boris: SaaS apocalypse is my favorite question. I think two things will happen—neither is what people are usually discussing.
First, anyone know the Acquired podcast? Yes, it’s a favorite of mine. Just recorded an unplugged episode with them—felt like meeting idols.
They introduced the “seven moats” concept, from Hamilton Helmer’s book—commercial competition’s seven core advantages. I think AI will make some moats less important and others more important.
Moats weakened:Switching costs: With models, migrating between systems is easy.Process advantage: Companies whose moat is workflow/process—Claude is increasingly good at optimizing workflows, especially 4.7; give it a goal to iterate, it improves until finished. It’s the first to truly achieve this.
Moats that remain strong: network effects, economies of scale, exclusive resources—AI hasn’t changed these.
Second: the number of startups over the past decade—I think the next decade will increase 10x. Now, a micro-startup can build a product at big-company scale and compete directly—big companies need to revamp workflows, retrain staff, huge internal resistance. Starting from scratch, you can be AI-native from day one.
So now is the best time to start a company, the best time for startups; opportunities for disruption are everywhere. We can achieve great things.
Lauren: Thank you, Boris. Audience Q&A time.Chapter Eight: Audience Q&A
Audience (Dan): You said you developed a product six months ahead of PMF; now models are good enough—how much of Claude Code’s success is due to the model, and how much to product decisions and experience?
Boris: Both—a mix. Six months ago, it was about 50/50.
Dan: In two years?
Boris: Two years... We plan in weekly increments, sometimes up to six months.
(Laughter)
By the way, the reason it was 50/50—years ago, I was at YC, did many startups, first employee at a YC company. YC emphasizes one thing: make something people love. However strong your model, you must make something people truly like. That’s why product detail matters—we obsess, so users have a great experience all day long.
As models get stronger, tool “scaffolding” becomes less important. Now we’re thinking: How to make loops first-class citizens? How to scale up agent parallelism? Sub-agents are just one idea; we’re doing much more.
I think in a year, models will align better, so today's mechanisms around prompt injection, static validation, permission modes, human review loops will become less important—models will just do things right.
Audience: I feel Claude Code drove a cultural shift, democratizing software development—now shopowners can develop software, or program microcontrollers for door lights. Will software development become like “using Microsoft Office”—a basic skill anyone can acquire, not just tech professionals?
Boris: Yes, absolutely! And I think it will go further—like “I know how to text.”
My reading preferences are mainly science fiction and tech history. In tech history, the closest analogy is the 15th-century European printing press.
Before the printing press, about 10% of Europeans were literate, employed for reading/writing by non-literate kings and nobles. After the printing press, in just 50 years, Europe published more texts than the previous thousand years combined, and book cost dropped 100-fold.
Full literacy took centuries, needing education systems and government support, but eventually global literacy reached about 70%. Now, we can all read/write without specialized degrees. Professional writing persists, of course.
I think software is undergoing similar democratization—at a pace much faster than 50 years. For example, writing accounting software—the best fit isn’t an engineer, but someone who excels in accounting, understands the field; writing code is the easy part. Domain expertise is the hard part.
Audience: Greg said your internal teams are “one step into the future” because you get early access to models and agents. Claude Code started as an internal tool, then launched externally. How far are you ahead of the outside world in engineering—one month, three, six? Is the gap widening or narrowing?
Boris: At the model level, we use the same as everyone else. Dogfooding is very important to us, so we use what others do—for example, some Mythos for testing, lots of Opus 4.7 for code. At the model level, there’s no real gap; future Mythos versions will become public.
The real gap is in the product—or more accurately, organizational processes. At Anthropic, we use Claude for almost everything, and our Claudes constantly talk to each other: My Claude runs loops, contacts others’ Claudes via Slack, runs loops to solve unknowns. No more handwritten code; all SQL is by models; everything is model-built.
So where we’re ahead isn’t technology—since everyone has the same tech. We’re essentially building a platform; developers and us use the same tools—the real lead is organizational structure and process. Through conversations like these, I hope everyone learns and evolves together.
Lauren: That’s also the startup advantage—it’s easier from scratch.
Audience (Jiren): We discussed multi-agent systems last time at Sequoia; lots still in the pipeline. Now there are slash/batch, loop, sub-agents, teams. Can you talk about how—model-wise or tool-wise—you inject prior knowledge, tweak objectives, so delegate/agent-launching is easier? Lots of work can be parallelized, but I still have to decide when; not yet model-driven.
Boris: Product-wise, it’s basically prompt tuning, that’s all. We tune prompts to get models to parallelize tasks more. Frankly, as models get stronger, they’ll just do so naturally. 4.7 now self-launches loops—e.g., I ask it for data queries, it notices the data changes over time and starts a loop to send me updates every 30 minutes. I just say yes, and ask if it can use Slack—it does so via Slack’s MCP.
So, over time, users shouldn’t have to figure out how to use tools better. If they do, it’s a product design issue—my failure. Models should handle this, with us guiding via prompt.
Audience: Many use Claude or Codex for heavy compute in the cloud, but some advocate “run AI locally.” As open-source models catch up, will this become mainstream in years? Will people stick to cloud or shift to local agent deployment?
Boris: I think the question itself may not matter—because models will fully handle code writing, agent launching, environment setup. If a local model is best, it’ll choose that. This won’t be an engineer’s concern anymore.
Audience (Jamie): Claude Code was successful partly by leveraging localized dev tools/workflow. For broader knowledge work, tools are often cloud-based. How does Claude deal with gaining tool access in these knowledge-work scenarios—can it be as powerful as for devs?
Boris: Moving big-company environments fully remote is a big job—I’ve done it, took five years. For knowledge work, most is cloud already: Salesforce, Google Docs, etc. For us, the simple answer is always: MCP. Connect Salesforce, Google Docs, Calendar via Claude’s MCP connector—Claude CLI, Code, anywhere can access.
Audience: For systems without MCP, is computer use a big opportunity?
Boris: Yes, computer use is a catch-all. Anthropic leads in computer use. With Claude, it can essentially operate any software on your computer—currently a bit slow, but especially after 4.7, it’s solid. Overall, MCP is top choice; whether MCP, CLI, API doesn’t matter to the model—it’s just tokens.Chapter Nine: Looking Ahead
Lauren: Last question.
Audience (Ryan): You said you anticipated product overhang and built ahead by six months, waiting for the models to mature. Can you describe, in abstract terms, what kind of product you’d build today to gain more value in six months or a year as models advance?
Boris: Claude Design is a good example. It’s already pretty good and will get better.
We have some features for Claude Code on the way, coming in the next few weeks—you’ll see soon.
Also, I think features like loop, batch—mass parallel agent execution—will keep evolving. Computer use is another good example.
Lauren: Boris, thanks so much for being here today. If anyone has more questions, Boris will be around—feel free to continue the conversation.
(Applause)
Boris: Thank you, everyone.
Risk Warning & DisclaimerThe market is risky, investments need caution. This article does not constitute personal investment advice, nor does it consider the special investment goals, financial situations, or needs of any individual user. Users should assess whether any opinions, viewpoints, or conclusions in this article are suitable to their own circumstances. Investment is at your own risk.