Claude Code Creator Reveals: Programming Has Been Solved, the Era of Universal Coding Is Here

Claude Code creator Boris Charney recently shared deep insights about the current state and future of AI programming at a developer event. As the central figure behind an internally incubated project at Anthropic, he not only recounted Claude Code's journey from obscurity to explosive growth but also made a bold prediction: programming is becoming a foundational skill like reading and writing.

From Accidental Birth to Product-Market Fit

Boris admitted that Claude Code's creation was largely serendipitous. In late 2024, he joined a small incubation team within Anthropic (Anthropic Labs). At that time, the state of the art in coding tools was still limited to "press Tab to autocomplete a line of code" inside an IDE — products like GitHub Copilot had already changed developers' daily workflows, but were fundamentally just extensions of autocomplete, far from reaching the ceiling of AI capabilities.

The team identified a clear "product overhang" — model capabilities far exceeding what existing products could deliver. This concept is gaining increasing importance in the AI industry: when the underlying model can already understand complex cross-file code logic and perform architecture-level refactoring, yet the product layer is still completing code line by line, that gap itself represents a massive opportunity. They decided to let the Agent write all the code directly, rather than continuing with line-by-line completion.

But reality was harsh: for the first six months, Claude Code was nearly unusable. Boris himself only used it for about 10% of his coding work. Even after public release, growth was flat. The real turning point came with the release of Opus 4 in May 2025, and every subsequent model iteration (4.5, 4.6, 4.7) brought exponential growth.

This reveals a key strategy: build products ahead of time for the next-generation model. The team clearly knew the current experience wasn't good enough, but they were betting on the future — and they won that bet. This is also the fundamental difference between AI application-layer startups and traditional software startups: you're not optimizing for current user experience, but building the product skeleton for model capabilities six months from now.

What "Programming Has Been Solved" Really Means

Boris conducted a live poll at the event: very few people write code entirely by hand, and not many write code entirely with AI either — most fall somewhere in between. But for him personally, the answer was clear — the model writes 100% of his code.

He revealed some staggering numbers:

He routinely commits dozens of PRs daily
One day last week he submitted 150 PRs (a personal record)
The Claude Code codebase itself was written with AI (TypeScript + React, chosen because they're widely distributed in model training data)

That last point is worth deeper reflection: when choosing a tech stack, the team no longer just considers language performance or ecosystem — they prioritize how proficient the AI model is with that language. TypeScript and React have massive volumes of open-source code on GitHub, meaning the model has seen vast amounts of high-quality TypeScript patterns and React best practices during training. This mindset of "optimizing tech stack choices for AI" could profoundly influence the technology landscape over the next few years.

Of course, he acknowledged that "programming has been solved" comes with caveats. For very complex large-scale codebases, or niche languages the model hasn't mastered yet, there's still a gap. But his advice is simple: just wait for the next model.

Boris's Extreme Workflow: Phone + Looping Agents

The most jaw-dropping revelation was Boris's personal development setup. Most of his work is now done through the Claude App on his phone:

Running 5-10 sessions simultaneously
Multiple Agents under each session
Hundreds of Agents running at any given time
Thousands of Agents executing deep tasks every night

His most praised feature is Loop — essentially having Claude automatically repeat work through Cron-scheduled tasks. Traditional Cron is a time-based scheduler born in Unix systems in the 1970s that executes fixed scripts at fixed times. Agent Loop is the AI-era evolution of this concept: instead of executing preset operations, it periodically wakes up an AI Agent with judgment capabilities, letting it make autonomous decisions based on the current environment. Boris's specific use cases include:

One Loop monitors PRs (fixing CI, auto-rebasing)
One Loop maintains codebase health (fixing flaky tests)
One Loop scrapes user feedback from Twitter every 30 minutes and clusters it

Boris emphasized: "I think Loop is the future." Anthropic also recently launched the Routines feature, migrating this capability from local to cloud — even when you close your laptop, server-side Loops keep running, delivering 24/7 uninterrupted Agent services. This fundamentally changes the traditional assumption that developers must be online to push work forward.

Future Team Structures: The Rise of Cross-Disciplinary Generalists

Regarding the future shape of teams, Boris gave a clear verdict: generalists will far outnumber today's count. But "generalist" here doesn't just mean full-stack engineers who can write iOS, Web, and backend code — it means cross-disciplinary polymaths who understand product engineering while also being proficient in design, data science, or user research.

He used the Claude Code team as an example: engineering managers, product managers, designers, data scientists, finance people, user researchers — everyone on the team writes code. This isn't a hard requirement but rather the natural outcome of AI lowering the barrier to programming. When writing code no longer requires years of specialized training, domain experts can directly translate their knowledge into software products without needing a "translation layer" — the traditional programmer role.

SaaS Doomsday? A Reshuffling of Competitive Moats

Facing the question "if AI reduces coding costs by 10-100x, will SaaS collapse?" Boris referenced Hamilton Helmer's "Seven Powers" framework for analysis. Helmer is a Stanford economics professor who, in his 2016 book 7 Powers, systematized seven sources of enduring competitive advantage: scale economies, network effects, counter-positioning, switching costs, branding, cornered resources, and process power. The book is deeply influential in Silicon Valley investment circles and serves as a standard tool for evaluating company moats.

Boris applied this framework to competitive analysis in the AI era:

Weakening moats:

Switching costs (models can help you migrate from one product to another — they can read the old system's data structures and automatically write migration scripts)
Process power (the 4.7 model can automatically optimize any process; give it an objective and it can hill-climb to the optimal solution — hill climbing here refers to the optimization approach of incrementally approaching the best solution; AI lets any company rapidly achieve process optimization)

Still-solid moats:

Network effects
Scale economies
Cornered resources

His more exciting prediction: the number of startups in the next 10 years will be 10x the previous decade. Small teams can build products of equal value to large companies, while big companies face massive inertia from internal resistance and process transformation.

The Printing Press Analogy: The Era of Universal Programming

When asked whether "programming will become as widespread as Office skills," Boris's answer was more radical than the questioner imagined — he believes programming will become as basic as "sending a text message."

He drew a historical analogy: the printing press revolution of the 15th century. Before the printing press, only 10% of Europe's population was literate — they were employed by illiterate kings and lords. Within 50 years of the printing press's invention, more literature was published than in the preceding thousand years combined, and book costs dropped 100x. While full adoption took 200 years, global literacy rates eventually rose above 70%.

The deeper meaning of this analogy: the printing press didn't just make "copying books" cheaper — it fundamentally changed the structure of knowledge dissemination and social power dynamics. Similarly, AI programming tools aren't just making "writing code" faster — they're changing the fundamental landscape of who can create software and who can solve technical problems.

"You don't need a degree in reading and writing to be literate, but professional writers still exist. Programming will be the same."

He specifically highlighted one insight: the best accounting software of the future won't be written by engineers — it'll be written by the best accountants. Because domain knowledge is the real challenge; coding itself has become the easy part.

Where the Real Gap Lies Inside Anthropic

An interesting question arose: is Anthropic significantly ahead of the outside world internally? Boris's answer was surprising:

No gap at the model level — they use the same model versions internally as externally (primarily Opus 4.7), because dogfooding is critical for platform products. Dogfooding is a classic tech industry practice requiring companies to be the first users of their own products to discover problems. For AI products this is especially crucial: if the developers themselves don't depend on their own tools, they can't truly understand users' pain points and workflows.

The real gap is in organizational processes:

All SQL within the company is written by models
No more hand-written code
Agents communicate with each other through Slack — when Boris's Claude is coding in a loop, it automatically contacts other colleagues' Claudes to resolve unknown issues

That last point is particularly noteworthy: it means AI Agents have already formed collaboration patterns similar to human teams. When one Agent encounters an uncertain problem, instead of stopping and waiting for human instructions, it proactively seeks help from other Agents through communication protocols. This autonomous inter-Agent collaboration may be the key to the next order-of-magnitude leap in organizational efficiency.

This means Anthropic's lead isn't in technology access, but in having already completed an AI-native transformation at the organizational level.

Key Trends to Watch

Boris concluded by mentioning several directions that are about to become significantly more powerful:

Claude Design — already decent, about to improve dramatically
Massively parallel Agents — Loop and Batch modes will become increasingly intelligent
Computer Use — as a complement to MCP (Model Context Protocol), handling software without APIs. MCP is an open protocol launched by Anthropic in late 2024, designed to standardize the communication interface between AI models and external tools, similar to how USB unifies hardware connections. But in reality, vast amounts of enterprise software lacks open APIs and doesn't support MCP. Computer Use solves this "last mile" automation problem by letting AI recognize screens and click buttons just like a human.
Increased model autonomy — future models will decide on their own when to launch Agents, whether to use local or cloud models; engineers will no longer need to make these decisions

His ultimate prediction: in a few more years, models will write all their own code, launch their own Agents, and build their own environments. The space for us as engineers to make decisions will shrink — and that's not a threat, it's liberation.