Full Breakdown
From Vibe Coding to Agentic Engineering: Andrej Karpathy’s Call for a New AI Development Paradigm
6/13/2026, 12:32:47 PM
Karpathy Declares Vibe Coding Obsolete
In February 2025, Andrej Karpathy introduced “vibe coding,” a workflow where developers type natural-language prompts and an AI model generates code. By February 2026, Karpathy announced that vibe coding is outdated and promoted “agentic engineering,” in which the model writes, tests, debugs, and iterates code autonomously.
Background and Early Adoption
Vibe coding emerged as the first wave of AI coding tools that could scaffold projects quickly, allowing developers to launch landing pages in an afternoon and prototypes over a weekend. Early adopters used it to accelerate shipping without expanding engineering headcount, mirroring the adoption pattern of Canva, which lets users create designs without specialized expertise.
Timeline of Key Milestones
Feb 2025: Karpathy coins “vibe coding.”
May 2025: He joins Anthropic to rebuild its pretraining team.
Early May 2026: Bloomberg video shows senior engineers using agentic engineering.
May 2026: PYMNTS reports a US neobank shipping a customer-facing feature built entirely by an AI agent.
Feb 2026: Karpathy publicly declares vibe coding obsolete.
Impact on Development Speed and Scope
Sources report that agentic engineering cuts feature-building cycles from a week to a day. PYMNTS noted a regulated banking feature that previously required four months of work was delivered in four hours. The same reports claim roadmaps can double in scope without adding team members, reshaping product planning and pricing.
Official Statements and Industry Observations
Karpathy argues that vibe coding was a bridge to a system where humans set goals and agents execute end-to-end development. Bloomberg’s senior engineers confirmed internal adoption of agentic workflows, noting they no longer use the term “vibe coding.” Harvard Gazette reported scientists, not professional coders, building functional internal tools within an hour using agentic engineering. PYMNTS highlighted a compliance team approving an AI-generated banking feature after a four-hour review.
Criticism, Conflicting Reports, and Gaps
Critics note that despite Karpathy’s claim, vibe coding still powers rapid prototypes in contexts where the full agentic workflow is not yet reliable. The banking case shows organizations continue to rely on the older workflow for speed. Sources provide no broad adoption metrics beyond the cited examples, leaving the overall diffusion of agentic engineering uncertain.
On-the-Ground Experiences
Engineers at three unnamed firms described pipelines where an AI agent receives a brief, iterates code, and returns a deployable artifact within an hour. Harvard researchers reported similar internal-tool builds, emphasizing the shift from coder to task director.
What’s Next
Karpathy’s move to Anthropic positions the company to shape future pretraining models that support agentic engineering. Observers expect additional case studies in finance, healthcare, and enterprise software, while regulators are likely to issue guidelines for AI-generated code in production.
