Karpathy's Software 3.0: From Vibe Coding to Agentic Engineering
> "I have never felt this behind." — Andrej Karpathy, Sequoia AI Ascent 2026
This Chinese forum post analyzes Andrej Karpathy's Software 3.0 thesis from his AI Ascent 2026 talk, arguing that the default workflow of programming has fundamentally changed.
From Vibe Coding to a New Paradigm
In 2025 Karpathy coined vibe coding: describing what you want in natural language while AI handles implementation. A year later at AI Ascent 2026, he reframed it: "Vibe coding only raised the floor. The real battlefield is elsewhere."
Three Paradigm Shifts
- Software 1.0: humans write explicit rules (if/else, loops).
- Software 2.0 (Karpathy, 2017): humans curate data and train neural networks; the program lives in weights (e.g., Tesla's vision stack).
- Software 3.0: humans program via prompts, context, tool use, examples, memory, and instructions. The LLM is an interpreter; "the context window is the new program." A Markdown install guide handed to an agent replaces a brittle shell script—precise-but-fragile code gives way to fuzzy-but-adaptive agent programs.
- Vibe coding drops the barrier to near zero but ignores quality, security, maintainability, and taste—fine for prototypes, dangerous for production.
- Agentic engineering raises the ceiling: agents generate, code, test, and check each other while humans write specs, review diffs, manage permissions, and preserve taste. Example: a MenuGen payment bug where an agent matched Stripe emails to Google login emails—plausible code, wrong system design (persistent user IDs were correct).
- App collapse: some categories (e.g., MenuGen, a menu generator) shrink from a full stack to a single model call—"Some app categories are collapsing into a single model call."
- Hiring: LeetCode is mismatched; instead "build a substantial project with agents, deploy it, make it secure, and have adversarial agents try to break it." The 10x engineer gap may become far more extreme.
- Understanding is non-outsourceable: "You can outsource your thinking, but never your understanding." Learn deep concepts, not API trivia.
- Compute inversion: GPU wrapping CPU; AI-native operating systems are already emerging (Copilot in Windows, Apple Intelligence).
- Karpathy, Software 2.0 (2017): https://karpathy.medium.com/software-2-0-a64152b37c35
- AI Ascent 2026 full video: https://www.youtube.com/watch?v=96jN2OCOfLs
- Karpathy tweet summary: https://x.com/karpathy/status/2049903821095354523
- Stephanie Zhan tweet summary: https://x.com/stephzhan/status/2049518659513852109
- Sequoia Inference analysis: https://inferencebysequoia.substack.com/p/andrej-karpathys-software-30-and
- Karpathy Bear Blog notes: https://karpathy.bearblog.dev/sequoia-ascent-2026/
- Anthropic, Effective Harnesses for Long-Running Agents: https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
The December 2025 Turning Point
Karpathy describes a personal threshold: during a December 2025 break, agent-generated code became larger, more coherent, and reliable enough that he stopped remembering when he last corrected it. Not that coding got harder—the default workflow changed.
The Unit of Programming Changed
Programming moved from typing code lines to delegating macro-actions: "implement this feature," "refactor this subsystem," "research this library," "write tests, run tests, fix failures." Humans become agent orchestrators. Karpathy jokes he developed "AI psychosis"—16 hours a day talking to agents.
Floor vs. Ceiling
> "The agent can remember whether a tensor library uses dim, axis, keepdim... The human still needs to understand the underlying concepts: storage, views, memory copies, invariants, identity, security boundaries."
LLMs Are Ghosts, Not Animals
LLMs are jagged, stochastic, summoned entities—brilliant at some tasks, inexplicably failing at others. This demands taste, judgment, and oversight as core skills.
Verifiability Framework
AI lands fastest where results are highly verifiable: code (compile, test) and math proofs lead; creative writing and medical diagnosis lag. Advice: seek "valuable but overlooked, highly verifiable domains."
Other Key Points
Closing Thought
When intelligence becomes cheap, understanding becomes scarce. The post updates Feynman: "What I cannot verify, I do not understand." Agents can create; verification and understanding remain human responsibilities.