Andrej Karpathy: From Vibe Coding to Agentic Engineering

Andrej Karpathy discusses the transition from "vibe coding"—using LLMs to quickly generate code—to "agentic engineering," where developers architect robust, autonomous systems. He argues that we are entering a "Software 3.0" paradigm where neural networks act as computers, shifting the developer’s role from writing syntax to designing prompts, managing context, and overseeing quality.

Chapters

Chapter 1: Introduction

• Andrej Karpathy, a founding member of OpenAI and former Tesla Autopilot lead, is introduced.
• He is credited with making complex technical shifts accessible and coining the term "vibe coding."

Chapter 2: Feeling Behind as a Coder

• Karpathy describes a shift in December where AI models became reliable enough to trust for full tasks without correction.
• This transition turned "vibe coding" into a coherent, agentic workflow that fundamentally changed his personal productivity.

Chapter 3: Software 3.0 Explained

• Software 1.0 is explicit code; Software 2.0 is learned weights; Software 3.0 is using LLMs as a programmable computer.
• The "context window" is the new lever, and the LLM acts as the interpreter performing computation in digital information space.

Chapter 4: Agents as the Installer

• Installation processes (e.g., OpenClaw) move from complex, brittle shell scripts to simple, high-level natural language instructions.
• Agents act as "installers," using their own intelligence to interpret the environment and debug issues in real-time.

Chapter 5: Menu Gen vs Raw Prompts

• Traditional app development often re-implements functionality that LLMs can now perform natively.
• Karpathy argues that we should move away from building apps that act as "wrappers" and toward direct information processing that was previously impossible.

Chapter 6: What’s Obvious by 2026

• The future may involve "neural computers" where neural networks are the host process and CPUs function only as co-processors.
• Deterministic tasks will continue to exist, but the heavy lifting of digital computation will be dominated by networked AI.

Chapter 7: Verifiability and Jagged Skills

• AI capability is "jagged"—it peaks in domains with clear verification (code, math) and struggles elsewhere.
• Models remain inconsistent ("jagged") because they are essentially reinforcement learning engines; if a domain isn't in the training mix, the model will fail.

Chapter 8: Founder Advice and Automation

• Founders should focus on verifiable domains where they can build their own reinforcement learning environments to fine-tune models.
• Everything is technically automatable if it can be broken down into verifiable outputs, even creative tasks.

Chapter 9: From Vibe Coding to Agent Engineering

• Vibe coding raises the floor for everyone; agentic engineering preserves the professional quality bar while moving much faster.
• Hiring should shift from solving isolated puzzles to evaluating a candidate's ability to build secure, complex projects and manage agent networks.

Chapter 10: Agents Everywhere and Learning

• We need "agent-native" infrastructure—tools and documentation designed for machines, not just humans.
• While you can outsource thinking, you cannot outsource understanding; human taste, judgment, and oversight remain essential "bottlenecks."