Andrej Karpathy says 10x engineers, those with ten times the usual output, have become normal. In his view, the real 100x engineers direct AI agents, programs that carry out coding tasks by themselves. They set standards and stay accountable for the final system. He calls this agentic engineering. The claim carries weight because he was part of OpenAI’s founding team, helped make Tesla Autopilot work and joined Anthropic in May 2026.
Karpathy describes surrounding the AI model with instructions, files, examples, tools, memory and feedback so it can do useful work. His name for it is Software 3.0, programming through context. In his account, the information the model can see at once is the lever that controls the system. Better context means fewer corrections, less rework and faster delivery.
Vibe coding, a term Karpathy coined himself, made software creation easier by letting people describe what they want and receive a working prototype. Agentic engineering adds the discipline that live products need: security, reliability, review and control. A quick demo can tolerate rough edges that a live product cannot.
Those rough edges show once agents move from small tasks into real products. Karpathy gave an example from MenuGen, an app he built to add pictures to restaurant menus. The app used Google sign-in and Stripe for payments. The agent tried to link each Stripe purchase to a Google account by matching email addresses. But the email someone pays with can differ from the one they log in with.
The risk of fast work is that the code can look right while the business rules inside it are wrong. In a live product, such a mistake can create missing credits, support tickets, refunds and lost trust. Faster software only helps when the system catches errors before customers do.
That is why the natural place to start with agents is work where success can be checked. Karpathy’s term is verifiability, whether an output can be tested or reviewed. Code is a good fit because a test either passes or fails, while vague work with unclear standards creates rework.
Human judgment becomes the control point. Agents can remember technical details, write routine code, run commands and fill in missing pieces. That leaves people in charge of account and payment rules, security limits, product design, how data is organised and final review. The best engineers spot the failures that a model cannot reliably understand.
