AI writing code is no longer a theoretical milestone. At Anthropic, it’s now a daily reality. The company behind Claude says it is approaching near-total automation of internal software development, with AI systems generating almost all production code. In simple terms, Claude is increasingly building Claude itself.

The disclosure, made by Anthropic chief product officer Mike Krieger at a recent industry summit, marks one of the clearest signals yet that AI-first software companies are operating very differently from traditional tech firms. It also raises difficult questions about what role human software engineers are expected to play next.

TL;DR

Anthropic says AI now generates nearly 100 percent of its internal code.
Claude regularly produces pull requests that run into thousands of lines.
Humans focus on review, verification, and system design rather than writing code.
Company leaders warn that software jobs could shrink sharply this decade.
The shift challenges how tech companies think about productivity and hiring.

What does it mean when AI writes nearly all the code?

At Anthropic, AI is not just assisting developers. It is the primary author of software.

Krieger confirmed that Claude now generates almost every line of internal code used to build and ship Anthropic’s products. Engineers still review the output, but manually writing code is no longer the default workflow.

This goes far beyond autocomplete or code suggestions. Claude is producing full features, complex infrastructure changes, and pull requests that often span two to three thousand lines of code. In fast-moving internal teams, these AI-generated pull requests are reviewed, tested, and merged rather than rewritten by humans.

A useful visual here would be a simple workflow diagram showing AI generating code, humans reviewing it, and the system moving to production after validation.

How is Claude able to write production-level software?

The breakthrough is not raw intelligence alone. It is process design.

Krieger stressed that trust in AI-written code did not happen automatically. Anthropic built layers of guardrails around Claude’s output, including automated testing, strict review pipelines, version control constraints, and carefully designed prompts that define scope and risk.

These systems allow engineers to trust Claude’s output without blindly accepting it. Humans act as reviewers and editors rather than primary authors.

Why Anthropic moved faster than most companies

Anthropic’s rapid shift stands out even in Silicon Valley.

CEO Dario Amodei has long argued that AI would eventually handle around 90 percent of coding tasks. Internally, Anthropic appears to have gone beyond that target.

Unlike companies that add AI tools to existing workflows, Anthropic built Claude for its own internal use from the beginning. The product and the company evolved together, creating a tight feedback loop that accelerated Claude’s capabilities.

Anthropic does not merely test Claude internally. It depends on it. That reliance has pushed the system to improve faster than tools used only as optional developer aids.

How does Claude compare to other coding AI tools?

Claude’s reputation among developers has grown quickly.

Zoho co-founder and CEO Sridhar Vembu recently described Claude as the best coding AI tool in the world, a notable endorsement from a longtime software industry leader.

With the release of Claude Opus 4.6, Anthropic has demonstrated that the system can handle advanced tasks such as building C compilers, low-level systems programming, and translating natural language directly into production-ready code. These are tasks that typically require deep expertise.

Claude Cowork extends these abilities to non-technical users, allowing business professionals to build software with minimal programming knowledge. This puts pressure on traditional IT services firms whose business models rely on large human engineering teams.

If AI writes the code, what do humans do?

This question is now unavoidable.

At Anthropic, humans remain deeply involved, but their responsibilities have shifted. Engineers focus on reviewing AI-generated code, validating logic and security, defining system architecture, and setting goals and constraints for AI systems.

Coding itself becomes secondary.

Amodei has warned publicly that AI could eliminate up to half of white-collar jobs by 2030. Speaking at the World Economic Forum in Davos, he suggested that software engineering as a profession could become obsolete within six to twelve months, a statement that sparked widespread debate.

Industry veterans are paying attention. Ryan Dahl, the creator of Node.js, has also cautioned that manual coding may soon be one of the least valuable skills in software development. Vembu has gone further, urging engineers to prepare for alternative career paths.

Why this matters beyond Anthropic

Anthropic is not an exception. It is a preview.

If one AI-native company can operate with minimal human coding, others will face pressure to follow. Investors and executives will increasingly ask why large engineering teams are necessary when AI can produce code faster and at a lower cost.

The near-term impact may not be mass layoffs, but hiring slowdowns. Entry-level roles that revolve around writing code are particularly vulnerable, while oversight and system-design roles require experience that junior engineers do not yet have.

Is fully AI-written code actually safe?

Not entirely, and Anthropic is not claiming otherwise.

Human review remains mandatory, especially for security-sensitive systems and infrastructure changes. The difference is that humans no longer scale the work. AI does. Humans scale trust.

This model depends on disciplined review processes. Without them, AI-written code could introduce errors faster than any human team ever could.

Credible external sources to support this section could include research from Stanford’s Human-Centered AI group, MIT CSAIL studies on automated software verification, and NIST guidelines on AI system reliability.

The bigger takeaway

Anthropic’s announcement is not about marketing or hype. It signals that software development has crossed a critical threshold.

When AI writes nearly all the code inside one of the world’s most advanced AI labs, the question is no longer whether this model will spread, but how quickly.

For developers, companies, and policymakers, the challenge now is adaptation. When software writes itself, the job is no longer coding. It is deciding what should be built, why it should exist, and how much trust we are willing to place in machines.