Andrej Karpathy: Software 3.0, Agentic Engineering, and Why You Can’t Outsource Understanding
At the recent AI Ascent 2026 conference, Andrej Karpathy—former Tesla Autopilot director, OpenAI co-founder, and founder of Eureka Labs—delivered a compelling keynote exploring how the role of the software engineer is being fundamentally rewritten by artificial intelligence.
In a wide-ranging discussion with Sequoia Capital’s Stephanie Zhan, Karpathy introduced key conceptual frameworks to understand the shift from traditional programming to AI-driven system orchestration.
From ‘Vibe Coding’ to ‘Agentic Engineering’
Karpathy drew a sharp distinction between two modes of building that have emerged with the proliferation of code-generation models:
- Vibe Coding: This describes the process of building software where the human developer describes a feature, and the LLM writes all the code. The developer simply runs it, observes if it works, and iterates through chat prompts. While “vibe coding” is incredibly powerful for prototyping, raising the floor of who can build software, it often lacks the architectural rigor required for complex projects.
- Agentic Engineering: This is the professional evolution of vibe coding. It involves managing teams of AI agents that write, test, debug, and review code under strict, automated guardrails. Agentic engineering focuses on verification, safety, performance benchmarks, and long-term system architecture.
“Vibe coding is fun and fast, but it is not how you build a reliable, secure enterprise system,” Karpathy explained. “For that, we need a disciplined transition to agentic engineering, where human engineers act as directors and verifiers.”
The Shift to Software 3.0
Karpathy expanded on his famous “Software 2.0” thesis (which argued that neural networks trained on data represent a new programming paradigm) by introducing Software 3.0.
Under Software 3.0, programming transitions almost entirely into natural language prompting and steering. The LLM functions as an interpreter that executes computational steps within digital environments. Instead of writing instructions, developers write specifications, evaluate outputs, and design the workflows that coordinate multiple models to perform complex tasks.
‘You Can’t Outsource Understanding’
Despite the automation of code generation, Karpathy issued a strong warning to current and aspiring software developers: you can outsource your code generation, but you can never outsource your understanding of the system.
“If you don’t understand how the underlying code works, how the database is structured, or how the networking protocol behaves, you will eventually get stuck,” Karpathy warned. “When the AI agent makes a subtle mistake or enters an infinite loop, only the human with deep, fundamental engineering knowledge can diagnose and resolve the issue.”
He urged engineers to focus on cultivating strong taste in software architecture, rigorous verification methodologies, and a deep understanding of computer science fundamentals. The future engineer will not be valued for their typing speed or syntax recall, but for their ability to verify correctness and orchestrate complex autonomous agents.
LLMs as ‘Ghosts’
In a memorable metaphor, Karpathy suggested that developers should stop thinking of LLMs as intelligent animals or human-like minds, and instead view them as “summoned entities” or “ghosts.”
“They are jagged, statistical entities,” he noted. “They have incredible knowledge in some areas, and bizarre, simple failures in others.” Directing them effectively requires a new kind of engineering discipline—one that values clear constraints, explicit expectations, and programmatic verification above all else.
Based on Andrej Karpathy’s presentation at AI Ascent 2026.