Former Tesla AI chief Andrej Karpathy now codes "mostly in English" just three months after calling AI agents useless

In October 2025, AI developer Andrej Karpathy dismissed AI agents, saying "they just don't work." Three months later, he's reversed course.

In a detailed post on X, former OpenAI and Tesla AI researcher Andrej Karpathy describes a fundamental shift in how he works. In just a few weeks, he went from about 80 percent manual coding and 20 percent agent usage in November to the exact opposite in December: 80 percent agent, 20 percent manual edits.

"I really am mostly programming in English now, a bit sheepishly telling the LLM what code to write... in words," Karpathy writes. While this "hurts the ego a bit," the ability to edit software in large "code actions" is simply too useful to ignore.

For Karpathy, this represents the biggest change to his coding workflow in roughly two decades of programming. He suspects something similar is happening right now to a double-digit percentage of engineers, while awareness among the public remains in the low single digits.

AI models still make mistakes like "a slightly sloppy, hasty junior dev"

Despite his enthusiasm, Karpathy warns against unrealistic expectations. The models still make mistakes, and anyone working with code that matters should "watch them like a hawk." The nature of errors has changed, though. They're no longer simple syntax errors but "subtle conceptual errors that a slightly sloppy, hasty junior dev might do."

Karpathy lists specific weaknesses: Models make wrong assumptions and keep building on them without checking. They don't manage their confusion, don't seek clarifications, don't surface inconsistencies, don't present tradeoffs, don't push back when they should, and they're also "still a little too sycophantic."

He's particularly critical of their tendency to overcomplicate things: the models bloat abstractions, don't clean up dead code, and implement inefficient constructions spanning 1,000 lines that could be reduced to 100 as soon as you ask.

2026 could bring a "slopacolypse" of low-quality AI-generated content

For 2026, Karpathy predicts a "slopacolypse" across GitHub, Substack, arXiv, X, Instagram, and digital media in general: masses of "almost right, but not quite" code that generally tends to work but is often low quality. Alongside real improvements, there will also be plenty of "AI hype productivity theater," Karpathy adds.

Still, Karpathy calls the shift a "net huge improvement," and going back to manual coding is hard to imagine at this point. He's particularly impressed by the agents' persistence—they never get tired or demoralized and just keep plugging away at problems where a human would have given up long ago.

The main effect for him personally isn't working faster but being able to do more things. He can now tackle projects that wouldn't have been worth the effort before and work on code that would have been impossible due to knowledge gaps.

At the same time, he warns that LLM coding will split engineers into two camps: those who primarily enjoy writing code and those who prefer building things. His own ability to write code manually is slowly atrophying, Karpathy admits. But being good at reviewing code doesn't mean you can still write it yourself, since generation and discrimination are different abilities in the brain.

Karpathy's dramatic reversal from his October skepticism

Tesla's former AI lead concludes that the capabilities of LLM agents—particularly Claude and Codex—crossed "some kind of threshold of coherence" in December 2025, triggering a phase change in software engineering. The intelligence part suddenly felt much more advanced than everything else: integrations, tools, knowledge, new organizational workflows, and processes. 2026 will be a "high-energy year" for the industry, Karpathy predicts.

As recently as October 2025, Karpathy held a completely different view. In an interview with podcaster Dwarkesh Patel, he argued we shouldn't talk about a "year of agents" but more realistically a "decade of agents." He called the shortcomings of agentic AI systems profound: the models were cognitively inadequate, not multimodal enough, had no functioning memory, and couldn't reliably handle complex computer tasks. His verdict back then: "They just don't work."

That he's now completely reversing course highlights just how quickly the technology has advanced in recent months.

Karpathy isn't alone in his assessment. Recently, an OpenAI developer said he no longer writes code by hand, predicting that companies will soon lose track of their own codebases. A Google engineer has also reported major productivity gains from agent-based coding.