Andrej Karpathy has a habit of naming things before the industry knows it needs a name. In February 2025, he coined “vibe coding,” the practice of building software by describing what you want and accepting what the model produces. One year later, at Sequoia Capital’s AI Ascent 2026 event, he declared it already obsolete.
The replacement: Agentic Engineering.
Karpathy pinpoints December 2025 as the inflection point. In November, he was writing roughly 80% of his own code, using AI for the rest. By December, that ratio had inverted. He was delegating 80% to agents. “I can’t remember the last time I corrected it,” he said during the fireside chat with Sequoia partner Stephanie Zhan. “I just trusted the system more and more.”
This was not a gradual improvement. It was a threshold crossed. Coding agents stopped being helpful autocomplete and started producing large, correct chunks of code that required no correction. His side projects folder exploded.
Software 3.0: The Context Window Is the New Code
To frame the shift, Karpathy extended his Software 1.0 / 2.0 framework into a third era. Software 1.0 is explicit code written by humans. Software 2.0 is programming by training neural networks on data. Software 3.0 is programming through prompts, where the context window becomes the lever and the LLM is the interpreter executing your intent.
“Your programming now turns to prompting,” he said, “and what’s in the context window is your lever over the interpreter that is the LLM.”
The practical implication is radical. Most existing software was built for humans to click through. In Software 3.0, the interface layer becomes irrelevant. Karpathy illustrated this with a personal story: he built an app called MenuGen that OCR’d restaurant menus and generated food images. He later watched someone feed a menu photo directly to Gemini, which overlaid images onto the pixels without any app at all. “This blew my mind,” he said. “That app shouldn’t exist.”
Vibe Coding vs. Agentic Engineering
Karpathy is precise about the distinction. Vibe coding raised the floor. Anyone can now build working software by describing what they want. That democratization is real and valuable. But vibe coding has a failure mode at scale: no oversight, accumulating technical debt, security vulnerabilities introduced silently.
Andrej Karpathy Agentic Engineering is the professional response. “Agentic engineering is about preserving the quality bar of what existed before in professional software,” he said. “You are still responsible for your software just as before.” The human role shifts from writing code to directing agents, reviewing output, catching failure modes, and maintaining architectural judgment. The best agentic engineers will outperform the best traditional engineers by more than 10x, not because they type less, but because they understand more.
Also Read: Sam Altman Wants to Rethink the OS
The Jagged Intelligence Problem
One of Karpathy’s sharpest observations concerns the uneven capability of frontier models. He calls them “jagged entities,” capable of refactoring a 100,000-line codebase or finding zero-day vulnerabilities, yet unable to reason about walking 50 meters to a car wash. The jaggedness comes from where reinforcement learning reward signals have been concentrated: domains with verifiable outputs like math and code.
The implication for builders: if your use case falls within the RL training distribution, agents fly. If not, they struggle. Understanding where your domain sits on that spectrum is now a core engineering skill.
The Irreplaceable Human
Karpathy closed with the line that resonated most: “You can outsource your thinking, but you can’t outsource your understanding.” As agents handle more of the execution, the human bottleneck concentrates in taste, judgment, and the ability to define what is worth building. The engineers who survive this shift are not the ones writing the least code. They are the ones who understand the most about why software matters.
The profession is being refactored. December 2025 was when it started. Agentic Engineering is what comes next.
