Karpathy: 'Don't Learn to Code Apps'

- Andrej Karpathy, a founding member of OpenAI and former Director of AI at Tesla, recently launched an AI education company called Eureka Labs. His perspective is shaped by a career that includes leading Tesla's Autopilot computer vision team and creating a popular deep learning course at Stanford. - This isn't a long-held view for Karpathy; in October 2025 he publicly dismissed AI agents as tools that "just don't work." By December 2025, he reversed his position, stating that his workflow had shifted to 80% AI agent utilization, calling it the biggest change to his programming routine in two decades. - Karpathy has termed this new paradigm "agentic engineering," where developers orchestrate and oversee AI agents rather than writing most of the code directly. This moves the developer's role toward system architecture, prompt engineering, and quality assurance, managing AI collaborators that handle implementation details. - The tools enabling this shift vary in approach and cost; for instance, Devin by Cognition Labs operates as a more autonomous software engineer and costs around $500 per month. In contrast, Cursor is an AI-first code editor that integrates deeply into a developer's existing workflow for a $20 monthly fee. - The developer community's reaction, visible in discussions on Hacker News, is mixed. Some express fear of falling behind, while others argue this is a domain-specific shift, comparing it to previous platform changes like the rise of web applications, and believe it's practical to wait for the ecosystem to mature. - This trend extends beyond just code generation into "agent orchestration," using frameworks like CrewAI, LangGraph, and Microsoft's Agent Framework to coordinate multiple specialized AI agents to accomplish complex tasks. This allows for building systems where different agents handle distinct parts of a workflow, such as data analysis, customer inquiries, and scheduling.