We’re barely three years into the AI revolution, but something akin to recursive self-improvement in AI systems might already be here.
Boris Cherny, the creator of Claude Code at Anthropic, has confirmed what many in the industry had been building towards: the AI coding tool is now 100% written by itself. In a post on X , Cherny responded to a tweet noting Anthropic’s remarkable shipping velocity, writing simply: “Can confirm Claude Code is 100% written by Claude Code.” The post, made on March 7, 2026, has since drawn over 133,000 views.
From 80% to 100%: A Rapid Progression
This confirmation is the culmination of a trend that has been building for nearly a year. Back in May 2025, Cherny disclosed that roughly 80% of Claude Code’s codebase was already being written by Claude Code itself. “Probably near 80. Yeah, it’s very high,” he said at the time on the Latent Space podcast, while noting that human code review remained part of the process.
By December 2025, Cherny went further, revealing that he had not opened an IDE even once during an entire month. Every line of code he shipped — 259 pull requests, 497 commits, 40,000 lines added, and 38,000 removed — was written entirely by Claude Code powered by Opus 4.5.
Then in January 2026, Cherny confirmed that Cowork, Anthropic’s new desktop automation product, was written “pretty much all” by Claude Code. The latest announcement of 100% is the logical endpoint of this trajectory.
Anthropic’s CPO Had Already Signaled This
Cherny’s confirmation follows remarks from Mike Krieger, Anthropic’s Chief Product Officer, who recently said that Claude products are “being entirely written by Claude”. Krieger had also recalled that Anthropic CEO Dario Amodei had predicted about a year earlier that 90% of code would be written by AI by end of 2024 — a prediction many dismissed as wildly optimistic at the time.
The reality, it turns out, has exceeded even that forecast. Anthropic’s internal teams are now regularly producing 2,000–3,000 line pull requests, with Claude handling the heavy lifting on code generation across the board.
What “100%” Actually Means
It is worth unpacking what full AI authorship means in practice. Cherny has been consistent in emphasizing that human oversight has not disappeared — it has transformed. As he explained in an earlier interview, “Someone has to prompt the Claudes, talk to customers, coordinate with other teams, decide what to build next.”
In a striking analogy, Cherny has previously described Claude as the horse and Claude Code as the harness: powerful on their own, but requiring a skilled human rider to be truly effective. The human is still in the loop — steering, reviewing, deciding what to build — but the act of writing code has been almost entirely delegated to AI.
This is not fully autonomous recursive self-improvement in the sci-fi sense. But it is something arguably more consequential in the near term: a production-grade AI system that meaningfully contributes to its own development, under human supervision, at a pace and scale that would be impossible any other way.
Engineering Is Changing, Not Disappearing
One of the more provocative implications of this milestone is its impact on software engineers. Cherny, however, has pushed back against the notion that developers are being made redundant. In a widely circulated response to questions about why Anthropic is still hiring over 100 engineers despite AI writing all its code, he argued that the role of the engineer is evolving, not disappearing.
Cherny has gone further still, suggesting that the title “software engineer” may eventually disappear altogether, replaced by roles like “builder” or “product manager” as coding becomes a universal skill rather than a specialized discipline. On his own team at Anthropic, he has noted that PMs, designers, finance staff — everyone — codes.
He has also shared a practical set of tips for using Claude Code that illustrate how deeply the tool has reshaped developer workflows: running multiple parallel Claude sessions, using plan mode for complex tasks, building reusable skill libraries, and letting Claude write its own operating rules via CLAUDE.md files.
A Milestone With Broader Implications
Claude Code was launched in February 2025 and has grown at a remarkable clip, generating over $500 million in annualized revenue and expanding usage more than tenfold in its first several months. It now accounts for roughly 4% of all public GitHub commits, with some projections putting that figure at 20% by end of 2026.
The milestone of 100% self-authorship, while still operating under human direction, is a significant inflection point. It suggests the feedback loop between AI capability and AI-assisted development has tightened considerably. Better models enable better tools; better tools accelerate the development of better models.
Whether or not this constitutes “recursive self-improvement” in a technical sense, the practical reality is striking: one of the most widely-used AI development tools in the world is now writing itself. And its creator is not alarmed — he is optimistic.
“Software engineering is changing, and we are entering a new period in coding history,” Cherny wrote in December. “And we’re still just getting started.”