Andrej Karpathy, co-founder of OpenAI and former head of AI at Tesla, published a post on X that sparked debate across the entire tech world. He did not write anything vague or speculative. He wrote specifically: AI coding agents essentially did not work before December 2025. Since December, they have. And that is, to put it mildly, quite a significant statement.
Karpathy is not someone who throws around words like that lightly. He helped build Autopilot at Tesla, was there at the founding of OpenAI, and his account on X is followed by hundreds of thousands of people in the industry. When he says that something has fundamentally changed, people pay attention.
So what exactly happened? The models gained significantly higher quality, long-term coherence, and persistence. They can work through complex, multi-hour tasks without losing the thread. Before December, they could not. Now they can.
A Weekend Project in 30 Minutes: Karpathy's Own Example
To show that this was not just an empty claim, Karpathy added a specific example from his own weekend. He wanted to build a local dashboard for analyzing video from his home security cameras. He gave the agent a single long instruction in English: log in to my DGX Spark, set up SSH keys, install vLLM, download and test the Qwen3-VL model, build a server endpoint for video processing, create a basic web interface, test everything, configure systemd services, and write me a final report.
The agent worked for approximately 30 minutes. It encountered several errors, searched for solutions online on its own, fixed them one by one, wrote the code, tested it, debugged it, configured the services, and returned with a completed report. In the meantime, Karpathy did not touch the keyboard. "Just three months ago, this could easily have been an entire weekend project," he wrote. "Today, you start it and forget about it for half an hour."
It is hard to communicate how much programming has changed due to AI in the last 2 months: not gradually and over time in the "progress as usual" way, but specifically this last December. There are a number of asterisks but imo coding agents basically didn’t work before December…
— Andrej Karpathy (@karpathy) February 25, 2026
Agentic Engineering: A New Name for a New Kind of Work
Karpathy has coined the term "agentic engineering" for this way of working. It aptly describes what is happening: developers are moving away from writing code line by line and instead launching AI agents, assigning them tasks in natural English, and reviewing their outputs in parallel.
He sees the greatest opportunity in what he calls "moving up the abstraction stack." This means building orchestrators—higher-level systems that manage multiple parallel AI agents at once, each with its own tools, memory, and instructions. It sounds complicated, but in practice it resembles management more than programming in the traditional sense.
When someone asked him whether hundreds of developers would be replaced by "a few skilled prompters," Karpathy rejected the idea. According to him, the word "prompter" distorts the whole concept. "At the highest level, deep technical expertise may be an even greater multiplier than before," he said. Tools do not replace knowledge. They amplify it.
Where It Does Not Work and Who Disagrees
It would be naive to think that everything is rosy. Karpathy himself acknowledged that agents work significantly better in some situations than in others. They perform best where the task is well specified and where the result can be verified or tested. Greenfield projects without legacy baggage are their natural environment.
Developer Rafał Kobyliński, who works with production code, directly challenged Karpathy. On user interfaces, networking, and concurrent processing—the things that actually break in production—he says he sees "barely better results than last year." Karpathy replied that perhaps he was "holding the tool wrong." Honestly, that sounds a little like the response of someone who knows the criticism contains a grain of truth.
Daniel Ost pointed out another problem: "When AI fails, debugging takes three times as long because you are trying to understand code you never wrote." And AI researcher Yacine Mahdid summed it up in one sentence: "You can delegate thinking, but you cannot delegate understanding."
Productivity Panic and the Impact on Junior Developers
Bloomberg reported on a "productivity panic" spreading through tech companies. Companies are frantically calculating whether AI tools such as Claude Code or Cursor are shortening development time or eliminating jobs outright. Or both.
Stanford research from 2025 showed that young workers in the software industry faced a 16% decline in job opportunities. And this is where the real problem lies. Karpathy says that expertise is becoming a greater multiplier. But what about a junior developer who is still building that expertise? The very activity through which they learned to delegate—writing code manually—is now itself being delegated.
The analogy with architects and CAD software holds: architects did not disappear when computers replaced manual drafting. Draftspeople did. Transitions never wait for those they displace.
Karpathy concluded: "This is definitely not business as usual in the software industry." And that is hard to argue with.
