Andrej Karpathy Retires Vibe Coding for Agentic Engineering

In December 2025, Andrej Karpathy said 80 percent of his code was generated by AI agents, a figure he described as arriving at what he called "the inflection point for agentic coding." It was less than a year after he coined the phrase "vibe coding" and within months he was declaring that concept passe. The 18-month arc, from casual AI prompting to what he now calls agentic engineering, is the clearest public account yet of where professional software development is heading.

The tweet that named a generation of programming

In February 2025, Karpathy posted on X that he had started to "fully give in to the vibes, embrace exponentials, and forget that the code even exists." The phrase named a practice that had no common label yet: accepting AI-generated code without close scrutiny, iterating by pasting error messages back into a chat window, building fast without studying the underlying implementation. He called it vibe coding. The term spread within days and had entered investor memos and style guides before the month ended.

This was not Karpathy's first time labelling a shift. In November 2017, he published "Software 2.0" on Medium, arguing that neural networks represented a new programming substrate: instead of writing explicit instructions, developers would define desired behavior and let optimization find the implementation. In September 2023, he extended the frame on X, describing large language models not as chatbots but as "the kernel process of a new Operating System", with the context window as RAM and agents as long-running processes. Each label arrived before the market had named the thing it described, and each one compressed the timeline to the next.

StartupHub.ai data shows 602 AI developer tools companies now tracked in our database, a category that barely existed in those terms when Karpathy wrote "Software 2.0." Of the 41 with disclosed funding stages, 20 (49 percent) remain at Seed, consistent with Karpathy's observation that the vibe-coding phase was real but unstructured: demand outpaced tooling discipline.

December 2025: the inflection point Karpathy named

The 80 percent figure was not a projection. Karpathy stated it as a description of his actual December 2025 workflow and later identified that month as a qualitative turning point for agentic coding. Agent-generated code chunks had become large enough, and coherent enough, that the practical question shifted from whether the agent could handle a task to how to verify its output at the scale he was now delegating. He noted at the time that he could not recall the last occasion when he had manually corrected the model's direction mid-task.

That shift defines what separates vibe coding from agentic engineering in Karpathy's framing. Vibe coding treats the agent's output as broadly acceptable and moves on: it is the approach that has lowered the barrier to prototyping and made AI-assisted development accessible to non-programmers. Agentic engineering starts from the premise that the agent is, as Karpathy put it at Sequoia Capital's AI Ascent event in April 2026, "fallible and stochastic," and requires professional oversight at every step. The tools can look identical. The discipline is not. By February 2026, coverage in The New Stack and elsewhere was already framing the shift as the successor to the casual phase Karpathy himself had named.

The Sequoia talk and the Anthropic move

At Sequoia's AI Ascent 2026 in April, Karpathy offered his most formal definition of agentic engineering. In the blog summary he posted afterward, he described it as "the professional discipline of coordinating fallible agents while preserving correctness, security, taste, and maintainability." He put the contrast directly: "Vibe coding raises the floor. Agentic engineering is about extrapolating the ceiling." On what the ceiling implies for individual productivity, he argued the 10x engineer benchmark was an underestimate: "10x is not the speedup people can gain. People who are very good at this can peak much higher than that."

He named four practices that define the professional path: writing detailed specs before prompting; reviewing agent-generated diffs critically for correctness and security; designing automated evaluation loops to catch regressions; and managing agent permissions across a codebase. Eval design, he said, is the skill most commonly underdeveloped even among experienced engineers moving to agentic workflows.

Six weeks after that talk, on May 19, 2026, Karpathy announced he was joining Anthropic to lead Claude pretraining research under Nick Joseph. His announcement on X drew 148,000 likes and more than 7,800 replies. His focus at Anthropic, the training phase that determines the core reasoning and knowledge of frontier models, is a structural extension of the Sequoia argument: the ceiling of what agents can deliver is largely set by the quality of their pretraining, which is now where he is working.

What it means

Karpathy's 18-month arc from "forget that the code even exists" to the "professional discipline of coordinating fallible agents" maps precisely to the phase the industry is moving through. The casual floor has been raised. The professional ceiling is now being defined by practitioners who understand both the capability and the limits of what current agents can reliably do. His position at Anthropic, working on the training that will determine what future agents can handle, suggests the ceiling he described at Sequoia is not a static target. Whether agentic engineering becomes a formal job title or simply what software engineering means by 2028 is the question his current work is positioned to answer.

Sources

• Andrej Karpathy on Medium: Software 2.0 (November 2017)
• Andrej Karpathy on X: LLM OS concept (September 2023)
• Wikipedia: Vibe coding
• Andrej Karpathy: Sequoia AI Ascent 2026 summary
• TechCrunch: Andrej Karpathy joins Anthropic's pre-training team (May 2026)
• The New Stack: Vibe coding is passe. Karpathy has a new name for the future of software.

Editorial standards: every claim is sourced. Tips: [email protected]