Andrej Karpathy on Vibe Coding: Notes for Agentic Engineers

Karpathy didn't disavow the term he coined — he reflected on it fondly ("good fun, and it almost worked") and drew a line between two different activities: casual "vibe coding," and the professional discipline he now calls "agentic engineering."

"Vibe coding" entered the vocabulary in early 2025 to describe a specific posture — what our research calls the "prompt and pray" approach: write a simple natural-language prompt, let the model generate the code, accept the output largely unchecked, and move on if it appears to run. Karpathy's own framing of the term was casual and exploratory — fitting for prototypes, hobby projects, and early experimentation where the cost of a mistake was low.

By early 2026, in his one-year retrospective, Karpathy's point was about maturation, not failure: as LLM agents became capable enough to be "a default workflow for professionals," the casual posture of vibe coding needed a more rigorous counterpart for production work — one that keeps the leverage of agents "without any compromise on the quality of the software."

The audience for this shift is broad: software engineers integrating AI into daily workflows, AI product leaders deciding how much autonomy to grant agent systems, security leaders inheriting the blast radius of unchecked generation, and enterprise decision-makers setting policy for AI-assisted development at scale.

• The shift reframes the human engineer's job from prompt author to director, architect, and validator.
• It reframes AI from a single-shot chatbot into networks of autonomous, coordinated agents.
• It reframes "done" from "the code runs" to "the code is programmatically verified."
• It reframes target scale from prototypes and MVPs to enterprise-ready production software.

None of this means AI-assisted coding is being abandoned — Karpathy's own framing treats agentic workflows as the evolution of LLM-assisted programming into a default professional workflow, not a retreat from it. The change is in how much oversight surrounds the generation step.

The through-line of Karpathy's framing: as you shift from writing code to orchestrating and overseeing agents, the professional's main currency is no longer syntax memorization or raw typing speed — it is verification.

The research offers a specific mental model for managing AI agents: treat them like fallible interns who have infinite speed and perfect memory but zero baseline judgment. That framing does real work. It explains why agentic engineering needs more structure than vibe coding, not less — an intern with infinite speed and zero judgment can produce enormous amounts of confidently wrong work before anyone notices.

Under this model, the engineer's job is threefold: define the boundaries an agent is permitted to operate within, establish rigorous automated testing suites that catch what the agent cannot judge for itself, and strategically orchestrate how multiple agents coordinate — all without giving up the speed advantage that made AI-assisted coding attractive in the first place.

This is a meaningful reframing of engineering value. It does not claim that writing code by hand becomes worthless, nor does it claim that AI agents will independently reach production quality. It claims something narrower and more defensible: the bottleneck moves from generation to verification, and the engineers who can build trustworthy verification scaffolding capture the value that used to sit with fast typists.

• Speed and memory are no longer differentiators — agents now have both in abundance.
• Judgment, boundary-setting, and test design become the scarce, valuable skills.
• Verifiability is positioned as a discipline to build, not a feature to buy off the shelf.

The research frames this as a genuine shift in what "quality" requires — maintaining it without compromising on speed, rather than trading one for the other.