Can an AI Really Code Itself? Inside Anthropic's Claude Code Phenomenon

Explore the bold claim that Claude Code wrote 80% of its own code. Understand what it really means, how the agentic AI tool works, and its implications for software development.

📅

✍️ Gianluca

Can an AI Really Code Itself? Inside Anthropic's Claude Code Phenomenon

An engineer at Anthropic recently made a bold claim: "Claude Code wrote 80% of its own code." That simple sentence has caused a stir in the tech world. If true, it suggests we may be entering a new chapter in software development, one where artificial intelligence is not just helping write code, but is deeply involved in building itself.

This idea grabs attention for a reason. For decades, software has been created by people, carefully typed line by line. Now, a tool that can understand, generate, and test its own codebase hints at something bigger. It is more than automation: it is a sign of AI becoming an active coding partner.

Key Questions:

  • What does "80% self-coding" really mean?
  • Is it full-blown autonomy, or clever use of automation?
  • Should developers feel excited or cautious?

Decoding the 80% Claim

The Podcast That Sparked the Buzz

This statement did not come out of nowhere. It came straight from Boris, the lead engineer behind Claude Code, during an interview on the Latent Space podcast. When asked how much of Claude Code was written by the AI itself, Boris answered: "About 80%." That number surprised both the host and many people who later heard the clip.

It is worth noting where this conversation happened. The Latent Space podcast is made specifically for AI engineers, not the general public. By choosing a technical platform for the reveal, Anthropic clearly wanted to speak directly to developers who could understand the context and limitations behind such a bold claim.

What Does "80% Self-Coding" Actually Mean?

Let's be clear: this was not a case of the AI running wild and building itself with no help. Boris made sure to add, "Humans still did the directing and definitely reviewed the code." That means people still decided what the software should do, gave instructions, and checked everything it produced.

The 80% figure likely refers to how much of the written code was generated by Claude Code after getting clear directions from human engineers. It does not mean the AI designed itself from scratch or made its own decisions about architecture or features.

Important: Anthropic has not explained exactly how they calculated that 80%. Was it based on lines of code, features completed, or something else? Without that detail, it is hard to compare this claim to other AI coding tools.

So, while the number is impressive, the truth is more grounded. Claude Code worked like a supercharged assistant: once humans defined the tasks, it helped write most of the code quickly and efficiently. But the big-picture thinking, the planning, and the decision-making? That still came from people.

How Claude Code Actually Works

Claude Code is not just another AI that spits out code when you ask. It acts more like a smart teammate sitting next to you in the terminal, reading your project, planning next steps, and carrying out tasks with surprising independence.

Agentic, Not Just Generative

Claude Code is described as "agentic", which means it can take action on its own: not just generate text, but also run commands, make changes, and interact with your development environment. It lives in the terminal, where developers work directly, and can scan, read, and edit huge codebases without needing to be spoon-fed context. Thanks to its "agentic search" features, it understands file structures, logic, and even dependencies in projects with millions of lines of code.

This makes Claude Code feel less like a code assistant and more like an intelligent command-line partner that knows what to do once you explain your goal.

Smart Features That Work Like a Developer

  • Multi-file Edits

    It can change several files at once during a refactor or when adding a new feature.

  • Test Support

    It writes tests, runs them, and even fixes ones that fail.

  • Git Integration

    Works with GitHub and GitLab. Can read issues, make commits, handle merge conflicts, and file pull requests.

  • Extended Thinking

    Developers can control how deeply it thinks, using commands like "think hard" or "ultrathink" for more thoughtful results.

Built on Unix Philosophy

Claude Code follows a "Unix utility" design approach. That means it is not a big all-in-one platform: it is a small, powerful tool that can be mixed into your existing workflow. You can use it however you like, with no need to change how you code.

Best Practices Encouraged by Anthropic:

  • Explore, Plan, Code, Commit: A clear workflow that keeps the process transparent
  • Granular thinking levels: Developers can pick how deeply the AI should think before acting
  • Test-driven development (TDD): Claude writes failing tests first, then fixes them
  • Use of subagents: For complex tasks, Claude can call on helper processes to focus on specific issues

Overview of Claude Code Features

CategoryClaude CodeClaude 3.7 Sonnet (Model)
TypeAgentic CLI toolFirst hybrid reasoning model on the market
Core FunctionalityCodebase analysis, multi-file editing, test generation, Git operationsPowers Claude Code; excels at coding and complex problem-solving
Key FeaturesAgentic search, terminal integration, configurable thinking modesVisible step-by-step thinking, customizable quality vs speed tradeoff
Design Philosophy"Unix utility": flexible, composable, extensibleOptimized for real-world tasks, state-of-the-art on SWE-bench
InteractionOperates directly in terminal; interacts with file system and GitVia API and tools like Claude Code; supports large context windows

Why Developers and Businesses Should Care

Claude Code represents a shift in how software gets made. For developers, it is a new kind of teammate. For businesses, it could reshape how teams work, cut costs, and open the door to bigger, faster ideas.

  • Faster Shipping

    Tasks that used to take hours or days can now be done in minutes. Developers spend less time on setup and more time on strategy.

  • Less Grunt Work

    Claude handles routine updates, renames functions across repos, refactors code, and patches test suites.

  • Bigger Ideas, Smaller Teams

    Features that once required full sprints might now be tackled by one or two developers guided by AI.

  • Broader Access

    The flexible prompt interface makes it accessible not just to seasoned developers but also to those with limited coding experience.

A Shifting Landscape

Claude Code is part of a broader trend in AI-powered development tools. Some are baked into IDEs like GitHub Copilot or Cursor. Others, like Devin or Cosine, aim to act more like AI teammates that deliver entire pull requests. Claude belongs to a group of command-line agents, favored for their power, flexibility, and ability to slot into custom workflows.

Among these, Claude Code is starting to stand out. Users have noted that its outputs often feel cleaner and more thoughtful than some of its peers. And its agentic design, combined with top-tier models like Claude 3.7 Sonnet, helps it perform smarter actions, not just spit out code.

The Big Caveats and Risks

Claude Code and tools like it are reshaping how software is built. But even as we celebrate its potential, we need to talk about what could go wrong. Behind every leap forward, there are new risks developers and businesses must carefully manage.

  • Quality Debt: Smart Does Not Mean Perfect

    AI can write a lot of code fast, but it does not always get it right. Claude Code, like other AI tools, learns from huge datasets filled with public code, some of it clean and efficient, but much of it messy, outdated, or buggy. That means it can repeat bad patterns or sneak in flaws that no one catches until later.

  • Security Holes: Hidden Dangers

    AI does not always understand the deeper risks in the code it writes. It might suggest libraries with known vulnerabilities or forget to properly secure sensitive data. If that code is not reviewed carefully, businesses could end up shipping products that are open to attack.

  • Black-Box Code: When AI Becomes Too Opaque

    One growing concern is that AI tools might start writing code that even expert developers cannot fully understand. This black-box behavior makes it harder to debug, maintain, or verify systems.

  • Skills at Risk

    As AI handles more of the routine coding, there is a danger that human developers may slowly lose touch with core programming skills. Over-reliance on AI could lead to less curiosity, weaker debugging abilities, or less hands-on understanding of architecture.

  • Who is Responsible? The Governance Puzzle

    When AI-generated code causes harm, who is to blame? Is it the company using the AI? The AI's creators? Or no one at all? Ethical and legal questions around accountability, copyright, and ownership are still unresolved.

Emerging Role: To handle these challenges, a new kind of role is emerging: AI code auditors. These professionals would review AI-generated code for security issues, bias, and reliability. They would need deep knowledge of both software engineering and how AI models behave.

Conclusion: A New Symbiosis, Not a Farewell to Humans

Claude Code's 80% self-coding claim marks a turning point in software development. It shows that AI can now do far more than autocomplete: it can read, plan, and build entire systems with human guidance. But this is not a story of machines replacing developers. It is about partnership.

Human insight still drives the vision. Developers define the goals, shape the architecture, and review every line that ships. Claude Code just speeds things up and takes care of the repetitive work.

To succeed with tools like this, teams need transparency, strong oversight, and a clear understanding of AI's limits. With that in place, the future of software looks faster, smarter, and more collaborative than ever.

The next generation of software is not just co-written with AI. It is co-run with it.

Resources and Links