Andrej Karpathy, one of the most influential figures in artificial intelligence and the creator of the term “vibe coding,” has declared that AI coding agents have reached a critical threshold that is fundamentally transforming software engineering. The former OpenAI founding member and Tesla AI director shared his observations in extensive notes posted on X (formerly Twitter) on Monday, January 26, 2026, detailing how his own coding workflow has undergone a dramatic transformation.
In his over 1,000-word reflection titled “random notes from Claude Coding,” Karpathy revealed that AI coding agents “crossed some kind of threshold of coherence around December 2025 and caused a phase shift in software engineering.” His personal coding workflow shifted dramatically in just one month: from 80% manual coding with autocomplete and 20% AI agents in November, to 80% AI agent coding and 20% manual edits and touchups by December.
Karpathy specifically highlighted improvements in Anthropic’s Claude Code and OpenAI’s Codex, with Claude Opus 4.5—released in late November—receiving particular praise from the engineering community. The AI pioneer admitted he is now “mostly programming in English now, a bit sheepishly telling the LLM what code to write… in words.” While this transition “hurts the ego,” Karpathy acknowledged it’s simply too powerful to ignore.
The shift comes with trade-offs. Karpathy noted he’s “already noticed that I am slowly starting to atrophy my ability to write code manually,” raising questions about the future of traditional coding skills learned in computer science programs. However, he also devoted an entire section to the “fun” he experiences while coding with large language models.
The post generated significant responses from engineers at leading AI companies. Boris Cherny, an Anthropic staffer and creator of Claude Code, revealed that his team is “mostly generalists” filled with 10x engineers, and “pretty much 100% of our code is written by Claude Code.” Cherny personally hasn’t made manual edits for over two months. Engineers from xAI, including Ethan He and Charles Weill, also chimed in, with Weill suggesting founders can now “divide themselves” with coding agents like VCs divide capital across portfolios.
Cherny acknowledged quality concerns with AI-written code, noting agents can overcomplicate solutions and leave dead code. His solution? Having AI review the AI-written code itself.
Key Quotes
AI coding agents crossed some kind of threshold of coherence around December 2025 and caused a phase shift in software engineering.
Andrej Karpathy, former OpenAI founding member and Tesla AI director, made this observation in his detailed notes about AI-assisted coding. This statement marks a pivotal moment in the industry, suggesting that AI coding tools have reached a level of capability that fundamentally changes how software is developed.
I really am mostly programming in English now, a bit sheepishly telling the LLM what code to write… in words.
Karpathy described his new coding workflow, highlighting the dramatic shift from traditional programming to natural language instructions. This admission from one of AI’s most respected figures illustrates how profoundly AI tools are changing the nature of software development work.
I’ve already noticed that I am slowly starting to atrophy my ability to write code manually.
Karpathy acknowledged a concerning side effect of relying heavily on AI coding agents. This observation raises important questions about the long-term implications for developer skills and what happens when traditional coding abilities decline across the industry.
Pretty much 100% of our code is written by Claude Code. For me personally it has been 100% for two+ months now, I don’t even make small edits by hand.
Boris Cherny, an Anthropic employee and creator of Claude Code, revealed the extent to which his team relies on AI coding agents. This statement from someone building AI coding tools demonstrates the confidence these developers have in their own products and suggests where the industry is heading.
Our Take
Karpathy’s observations represent more than just a workflow optimization—they signal an existential shift for the software engineering profession. The December 2025 “threshold” he identifies aligns with the release of Claude Opus 4.5, suggesting that model improvements have reached a tipping point where AI coding becomes not just helpful but dominant. What’s particularly striking is the speed of adoption: a complete workflow reversal in just one month among elite engineers. The fact that Anthropic’s own Claude Code team writes 100% of their code using AI creates a fascinating recursive loop—AI engineers using AI to build better AI coding tools. However, the acknowledged “atrophy” of manual coding skills should concern the industry. We’re potentially creating a generation of developers who excel at prompt engineering but struggle with fundamental programming concepts. The question becomes: what happens when AI tools fail or produce subtle bugs that require deep technical understanding to debug? This transition may be inevitable and beneficial, but it demands thoughtful consideration of how we train future engineers and maintain critical coding knowledge.
Why This Matters
This announcement from Andrej Karpathy carries enormous weight for the software engineering industry. As a founding member of OpenAI and former Tesla AI director, Karpathy’s observations signal a fundamental shift in how code is written and what skills developers need. The rapid flip from 20% to 80% AI-assisted coding in just one month suggests we’re at an inflection point where AI coding tools have achieved sufficient reliability and capability to become the primary development method.
The implications extend far beyond individual workflows. If leading engineers at companies like Anthropic are writing 100% of their code through AI agents, this represents a paradigm shift in software development that will impact hiring practices, education requirements, and the very definition of what it means to be a software engineer. The acknowledgment that traditional coding skills may “atrophy” raises critical questions about computer science education and career development. For businesses, this could mean dramatically accelerated development cycles and reduced engineering costs, but also new challenges around code quality, maintainability, and the need for different skill sets focused on prompt engineering and AI oversight rather than manual coding.