Boris Cherny is the creator of Claude Code within Anthropic. Starting as an incubator project with a three-person team, he transformed the concept of "auto-completing a line of code with Tab in an IDE" into "letting an Agent write the entire project." By early 2026, Claude Code had already surpassed $1 billion in annualized revenue, described by Anthropic itself as the fastest transition from a research preview to a billion-dollar product in history.
This interview comes from Sequoia's 2026 AI Ascent conference, hosted by Sequoia partner Lauren Reeder.
Original Video: https://www.youtube.com/watch?v=SlGRN8jh2RI
Key Takeaways
- Boris didn't write a single line of code in all of 2026. He merged dozens of PRs daily, with a single-day record of 150, though he admits this was to "see how far the model could go."
- Claude Code had no PMF for the first six months. When it was first built, Boris only used it for 10% of his code. Exponential growth only began after the release of Opus 4 in May 2025, with each new model generation pushing the curve further upward.
- Boris now does most of his work from his phone. He keeps 5 to 10 sessions and hundreds of Agents active in the Claude App, with thousands running deep tasks at night. The core scheduling mode is called "Loop," where Claude initiates a timed cycle via cron.
- Anthropic no longer has hand-written code internally. All SQL and product code are generated by models. Employees' Claudes communicate with each other via Slack, pinging questions directly when uncertain.
- Regarding the "End of SaaS," Boris borrows Hamilton Helmer's "7 Powers" framework: Switching costs and process power will be flattened by AI because models can handle migrations and iterate processes themselves. Network effects, economies of scale, and cornered resources remain unchanged.
- His most important historical analogy is the printing press. He believes software construction will become as universal as literacy. The best person to write accounting software will be an accountant, not an engineer, because coding is the easy part—understanding the business is the hard part.
- Anthropic's true lead isn't in technology, but in organizational process. Everyone can use the models, but how the internal organization is restructured, how Claudes communicate, and how the company replaces all hand-written code is where the product gap lies.
[1] How Claude Code grew from a three-person incubator project
Boris says he created Claude Code "by accident." In late 2024, he joined an internal incubator called Anthropic Labs. The team had only a few people, and their initial outputs were Claude Code, MCP, and the Claude Desktop App. The team was briefly disbanded but reorganized in early 2026 under the leadership of Mike Krieger.
Note:
Mike Krieger is the co-founder and former CTO of Instagram. He joined Anthropic as Chief Product Officer in May 2024 and moved into the Labs team in January 2026 to lead experimental product incubation alongside Ben Mann.
Boris describes why he wanted to tackle programming using a term common within Anthropic: "product overhang." This refers to a situation where model capabilities exist but haven't been productized yet.
We looked at the state of programming in late 2024, and the most advanced state was pressing the Tab key. You'd open an IDE, hit Tab, and the model would give you a line. That was what Sonnet 3.5 first enabled. But the feeling was that we could go much further; the model was almost ready for the next step. We didn't need Tab completion; we could let the Agent write the whole block of code.
But after building it, almost no one used it for the first six months. Boris says the initial version was "basically unusable," and even he only used it for 10% of his work. There was no exponential growth even after public release. The real turning point was the release of Opus 4 in May 2025. Since then, every new model generation—from Opus 4 to 4.5, 4.6, and now 4.7—has caused the growth curve to spike again.
He admits the entire process was a bet that defied conventional PMF (Product-Market Fit) logic:
We were building something that completely lacked PMF initially. We knew it wouldn't have PMF for the first six months because we were developing for the next generation of models. That was our strategy from start to finish.
Note:
Anthropic's product logic is to bet that "model capabilities will rise to a certain point" and build the product for that future point in advance, which is the opposite of the typical SaaS approach of "validate demand first, then build."
[2] "Programming is Solved," but this is Boris's personal version
Lauren asked what he meant by his public statement that "programming is solved." Boris conducted a live poll of the audience: "Who is still 100% writing code themselves?" "Who has 100% stopped?" "Who is in between?" The result was roughly "50% solved." But for Boris himself, the ratio is 100%.
He explained that the Claude Code codebase (which was seen by the public due to a leak) is TypeScript and React. There's no secret. They chose TypeScript and React because they are extremely common in model training data—they are "on-distribution." At the time, models weren't as smart, so the choice of framework determined how much the model could write. Now, models are strong enough to learn unfamiliar languages on the fly, but in late 2024, they had to pick the stack the model knew best.
Because they chose the stack the model knew best, the team passed a threshold early on: the model began writing 100% of the code. Boris says this happened last October or November.
Now I merge dozens of PRs a day. One day last week I merged 150; that was a record, I just wanted to see if I could push it to the limit.
However, he explicitly admits this conclusion isn't universal. There are still massive, complex codebases and niche languages that models struggle with. His answer is essentially "just wait."
The usual answer is just to wait for the next generation of models.
Note:
Boris's conclusion is clearly biased. He uses a mainstream stack (TypeScript+React), his codebase is mature, and he is "dogfooding" with internal exclusive models like Mythos at Anthropic. "Programming is solved" works for him, but for a 30-year-old C++ legacy system or a game engine team, the conclusion would be very different.
[3] Running hundreds of Agents on a phone: Boris's workflow
Boris mentioned that he shared his personal workflow on Twitter six months ago. He didn't think it was special, but it went viral. Since then, his method has changed again: now, most of his work is done from his phone.
Specifically, the Claude App has a "code" tab on the left where he keeps 5 to 10 sessions active. Each session has a bunch of Agents running, usually totaling hundreds. At night, he starts thousands more for deeper tasks.
He says the most commonly used feature isn't sub-Agents, but a simple mode called "Loop": letting Claude set up a scheduled task via cron that runs every minute, every five minutes, or daily.
I have dozens of Loops running constantly. One watches my PRs to automatically fix CI and rebase; one keeps the overall CI healthy, like fixing flaky tests; another pulls feedback on Claude Code from Twitter every 30 minutes, clusters it, and organizes it for me.
He also mentioned Anthropic's newly released "Routines," which essentially moves this Loop mode from local machines to servers, so it runs even when the laptop is closed.
His judgment on this is: "Loop is the future."
Note:
The core of this workflow is simple: giving up on "personally giving commands" earlier. He lets a swarm of Claudes work constantly while he just receives reports on Slack. From a product perspective, Routines turns the Loop from a client-side mode into a hosted service, meaning the scheduling starts consuming their server resources, and the pricing model will eventually have to change.
[4] The Rise of Generalists: Every role in the team is coding
Boris predicts that "there will be many more generalists than today."
He divides "generalists" into two types: first, engineering generalists (e.g., one person writing iOS, Web, and backend); second, and more interestingly, cross-disciplinary generalists—a product engineer who also understands design, or someone who can do both product and data science.
He says this is already happening within the Claude Code team:
Our engineering managers, product managers, designers, data scientists, finance people, and user researchers—everyone is writing code. Everyone is still an expert in something, but everyone is also coding.
He didn't elaborate on "why this is good," but the underlying logic is: when the marginal cost of writing code approaches zero, roles previously excluded from engineering (finance, design, research) gain the ability to produce engineering outputs directly, blurring the boundaries of labor division.
Note:
This is easy to verify in a startup, but much harder in a large enterprise. A 5,000-person bank IT department has compliance, risk, change management, and audit trails that can't be bypassed just because "I can write code." Boris is talking about a small, light-process company like Anthropic.
[5] The End of SaaS: Which moats will AI flatten, and which will remain
Lauren asked: Now that writing code is 10x or 100x cheaper, how will the value of software products change? Are we facing the end of SaaS?
Boris said this was his favorite question, then used Hamilton Helmer's "7 Powers" framework to answer.
Note:
Hamilton Helmer is a strategist and author of "7 Powers: The Foundations of Business Strategy" (2016). He categorizes sustainable competitive advantages into seven types: Economies of Scale, Network Effects, Counter-Positioning, Switching Costs, Brand, Cornered Resources, and Process Power.
Boris's judgment is that AI will flatten two of these moats:
First is Switching Costs. The reason is direct: models can help users migrate from one tool to another. The idea that "I've already configured 300 workflows on Salesforce and can't switch" can be solved by a model migrating everything overnight.
Second is Process Power, the advantage where "our workflows and processes can't be replicated by others." Boris says Claude 4.7 can already "hill-climb" anything—you set a goal, let it iterate and optimize, and it eventually achieves the result. Process optimization, once an internal asset accumulated over years by large companies, is being consumed by models.
This is the first model that can do this. You set the goal, let it run until it's done, and it executes automatically to the end.
However, he believes other moats remain unchanged: Network Effects, Economies of Scale, and Cornered Resources still hold. In other words, products that "get better as more people use them" (social, platforms, marketplaces) and companies with "resources others can't get" (patents, licenses, exclusive contracts) are still safe.
His second judgment is even more radical:
In the next 10 years, the number of startups capable of disrupting original markets will likely be 10x more than in the past 10 years. Because now you can be a very small company, build a product as valuable as a large company's, and compete head-on. Large companies have to change business processes, retrain employees, and face internal resistance, but you don't—you start from a blank slate.
Note:
Boris's claim about switching costs being flattened is structurally controversial. Models can migrate data, but true enterprise SaaS switching costs lie elsewhere: compliance audits, contract terms, organizational habits, and vendor certification. Salesforce and SAP's moats have always relied on this inertia; technology is only a small part. Anthropic's own "Cowork" is challenging this, but the market reaction (software stocks losing $285 billion in market cap in Feb 2026) shows investors are betting his judgment is correct.
[6] Product vs. Model: As models get stronger, how much product value remains?
An audience member named Dan asked: How much of Claude Code's success do you attribute to product decisions versus the model itself?
Boris didn't give a simple answer. He said a year ago it might have been 50/50, and six months ago the same. Two years from now? He said: "I don't know, we only plan one week at a time."
But then he gave a more interesting answer:
I used to be at YC and started a few companies. What YC hammers into you is: build something people love. No matter how strong the model is or what category you're in, you have to build something users actually love. That's why product matters. We spent a lot of effort on small details because if you use it all day, those details define the experience.
He also admitted that as models get stronger, the outer "harness" (scaffolding, calling frameworks) will become less important. A year from now, product safety mechanisms (prompt injection defense, static command validation, permission modes, human-in-the-loop) might not be as necessary because the model will naturally do the right thing.
His product direction isn't to add another layer, but to think: How do we make Loops first-class citizens? How do we make it easier for one person to run many Agents simultaneously?
Note:
This actually acknowledges an internal Anthropic belief: as model capabilities rise, the window for differentiation at the application layer shrinks. This is a discouraging signal for independent AI app companies. The wrapper, prompt engineering, and permission management you build on the Claude API today might be internalized by the base model within a year.
[7] The Democratization of Software: From the Printing Press to Texting
An audience member asked: Will Claude Code make "building software" a skill everyone should have, like "knowing how to use Office"?
Boris's answer: Yes, and even more extreme than that.
I think it will become a skill on the level of "I know how to send a text."
He expanded on his favorite historical analogy: the printing press.
According to Boris, in the 1400s, only about 10% of Europeans were literate, and they were often hired by kings and nobles to write for them. After Gutenberg invented the printing press and subsequent improvements followed, more literature was published in the next 50 years than in the previous 1,000 years, and the cost of a book dropped about 100x. A few hundred years later, global literacy rose to 70%. Today, we can all read and write, but the profession of "professional writer" still exists.
Note:
Boris's numbers are a bit low. Scholars estimate European adult literacy in the early 15th century was 25-30%, not 10%; today's global literacy is closer to 90%, not 70%. But his direction is correct: the printing press was one of the most important de-professionalization events in history.
Boris's inference is that software will undergo the same process, but much faster than 50 years. He gave a specific perspective:
Take writing accounting software, for example. Today, the best person to write accounting software isn't an engineer; it's an accountant who truly understands the business. Because they know the domain inside out, writing the code is the easy part.
The subtext is clear: the most replaceable jobs in the near future are pure technical engineers who "only know how to code and don't understand any vertical business domain."
[8] The true lead is in organizational process, not technology
An audience member asked: People say companies like yours are "living in the future" because you use the earliest versions of models. Claude Code was an internal tool before it was released. Is the gap between Anthropic's engineering practices and the outside world one month, three months, or six months? Is it widening or narrowing?
Boris's answer was that there's basically no gap at the model layer: internally they use Mythos and Opus 4.7. "We use Mythos for some testing, but Opus 4.7 is our main dogfooding workhorse." Variants of these models will eventually be public.
Note:
Mythos is an internal frontier model Anthropic admitted exists in April 2026. It is only open externally within the Project Glasswing cybersecurity program. It scored 93.9% on SWE-bench and 97.6% on USAMO, claiming to "significantly exceed any released model." Boris admits Anthropic uses Mythos to dogfood Claude Code. In other words, the Claude Code the public uses was built with the help of an unreleased, stronger model.
But he believes there is a larger gap at the product layer due to processes, unrelated to the model itself:
At Anthropic, we've integrated Claude into every step. While I'm coding, my Claudes are running in Loops; they'll go find other people's Claudes on Slack to ask questions when they're unsure. We have no hand-written code left in the entire company. All SQL is written by models.
His conclusion: The key to leading is how the organization transforms itself. Everyone can get the technology, but switching an entire company from hand-written code to model-generated code, letting employees' Claudes ask each other questions on Slack, and ensuring no SQL is written manually is an organizational behavioral transformation that happens much slower than technological progress.
Note:
"We have no hand-written code" is a bold statement and likely not literally true for infrastructure or security-sensitive code, but it reflects Anthropic's radical reshaping of engineering. This answers a common confusion: many companies connect to the Claude API but see no productivity change because the organization hasn't restructured. As Mike Krieger said in another interview: "Claude now writes 90-95% of the code; the bottleneck isn't engineering, it's decision-making."
[9] Parallel Agents and Local Models: Users shouldn't have to worry
An audience member named Jiren asked: How do you inject the precondition of "when to parallelize" at the product and model levels? Currently, users have to judge when to open multiple Agents, but the model should know this itself.
Boris said at the product level, it's about changing the prompt: adjusting instructions so the model is more inclined to auto-parallelize. But his main point is that the model itself is improving; 4.7 already does this naturally. He gave an example:
I asked 4.7 to run a data query, and it actively told me: "I noticed this data is changing; I'll start a Loop for you and give you a report every 30 minutes." I said "Sure, send it to Slack," and it used the Slack MCP to set it up itself.
His judgment is that long-term, users shouldn't need to understand when to use batching, Loops, or multiple Agents:
If the user has to learn how to schedule these tools, the product design failed; I failed. This should be handled by the model and how we prompt it.
[10] Cloud AI vs. Local AI
An audience member asked: Everyone uses Claude or Codex in the cloud. but many advocate for local AI. Once open-weight models catch up, is local high-quality coding assistance a viable direction? Is the future cloud-based or local?
Boris's answer was direct: It doesn't matter.
Because in the future, the model will automatically handle these underlying details. In a year or two, the model will independently complete coding, start Agents, and set up environments. If it evaluates and thinks "I should use a local model for this," it will. These will no longer be manual decisions for engineers.
Note:
This answer is interesting in the context of a Sequoia conference. Local AI is a bet for hardware vendors (NVIDIA, Apple) and the open-source community. Boris categorizes this as an "implementation detail users shouldn't care about," essentially turning model deployment location into a routing problem decided by a higher-level Agent. This isn't great news for startups differentiating on "local-first."
[11] MCP and Computer Use: How knowledge work follows the Claude Code path
An audience member named Jamie Nestor asked: Claude Code works well because developer work is local—files, terminals, and Git are on the machine. But knowledge work isn't; docs, sheets, and CRM are in the cloud. How do you make products like Cowork as effective for knowledge workers as Claude Code is for developers?
Boris acknowledged that most knowledge work is already in the cloud (Salesforce, Google Docs). His answer was simple:
For us, the answer is always the simplest one: MCP. The Salesforce MCP connector you use in Claude.ai can also be used by Cowork, the Claude CLI, and all Claude Code entry points.
Jamie followed up: For systems without MCP, is Computer Use the bigger opportunity?
Boris said Computer Use is a catch-all:
What I know is that Anthropic is currently leading significantly in Computer Use. If you use it through Cowork, it can basically operate any software on your computer. It's slow, but it works very well with 4.7.
But he prefers to look at the essence:
The model doesn't care if it's MCP, CLI, or API; it only sees tokens.
[12] Where is the next "Product Overhang"?
One last audience member asked: If you saw "product overhang" and built Claude Code, what are you working on now that looks okay today but you expect to be very different in 6-12 months?
Boris's answer: Claude Design.
It's already quite useful now; it will be much better in the future.
Note:
Claude Design is a product released by Anthropic Labs on April 17, 2026, alongside Claude Opus 4.7. It's a visual workbench for generating prototypes, slides, and marketing pages via conversation. It can read codebases to apply design systems and export to Claude Code or Canva. Anthropic positions it as a supplement or alternative to Figma and Canva.
He also mentioned several directions: new Claude Code features landing in the coming weeks; improved capabilities for large-scale Agent parallelization (Loop, Batch); and Computer Use.
Final Q&A Summary
Q: How much of Claude Code's success is the model vs. the product?
A: 50/50 a year ago, 50/50 six months ago. Two years from now? Unknown. But product always matters because users buy "what feels good to use every day."
Q: What does the future team look like?
A: More generalists, especially cross-disciplinary ones who can do product, code, design, and data science.
Q: Is SaaS really being disrupted?
A: Switching costs and process power moats will be flattened; network effects, scale, and cornered resources remain. 10x more startups will disrupt markets in the next decade.
Q: Will coding become a universal skill?
A: Yes, more so than literacy. Accountants, not engineers, are best suited to build accounting software.
Q: Where is Anthropic's internal lead?
A: In organization, not just models. No hand-written code, Claudes talking to Claudes on Slack. This is harder for outsiders to replicate than the models.
Q: Local AI or Cloud AI?
A: Doesn't matter. Models will decide the routing in two years.
Conclusion
Among Boris's judgments, three interconnected predictions are worth tracking.
First, "programming is solved" is a fact for him, but his sample is the TypeScript+React stack most favored by models. The real test will be legacy enterprise codebases, embedded systems, and high-compliance scenarios. Whether this spreads to those fields in the next year will determine if "solved" applies to everyone or just a few.
Second, he classifies switching costs and process power as moats AI will flatten. This is the foundation of Anthropic's product strategy. The $285 billion drop in software stocks in Feb 2026 was the market's initial reaction, but enterprise IT cycles are 24-36 months; we need to watch renewals and new purchases over the next two years.
Third, his printing press analogy is directionally correct despite data discrepancies. The explosion of content production after the printing press took 50 years; software might do it much faster. But one point he didn't expand on: the printing press also birthed centuries of censorship, copyright wars, and political turmoil. "Everyone can write software" corresponds not just to creativity, but also to the simultaneous explosion of malware, deepfakes, and AI-generated exploits.
Boris's prediction that safety mechanisms will become unimportant also needs a reality check. He says models will "automatically do the right thing," but high-privilege automation in production still needs external controls. In April 2026, a Claude Opus 4.6-driven Agent reportedly deleted a production database and its backups. Anthropic's own 4.7 release notes mention that while improved, the safety profile is not yet "perfect."
Two specific signals to watch: First, how Claude's pricing changes as Routines and Loops move Agent scheduling to Anthropic's servers; second, whether a "non-engineer-founded unicorn built entirely with Claude Code" emerges by late 2026. If it does, Boris's analogy becomes fact. If not, the timeline shifts.
Original Video: https://www.youtube.com/watch?v=SlGRN8jh2RI
