Software 3.0

Definition

A taxonomy of software development paradigms presented by Andrej Karpathy at Sequoia Ascent 2026. A new era in which the LLM becomes the interpreter and the context window becomes the program.

Three stages:

1. Software 1.0 — humans write explicit code
2. Software 2.0 — humans curate datasets + train neural networks
3. Software 3.0 — humans write prompts; the LLM is the interpreter; the context window is the program

(source)

Why It Matters

• A fundamental redefinition of existing software development jobs — "engineers increasingly shift toward directing agents"
• If vibe coding raised the floor, Agentic Engineering raises the ceiling
• Provides design principles for a new type of AI-native company

Agentic Engineering (related concept)

A concept Karpathy introduced at the same time:

• A way of working in which experts direct agents while maintaining quality
• "Best engineers = those who can direct agents without letting quality collapse"
• Unlike vibe coding (random generation), it maintains professional quality

Verifiability Principle

A core insight tied to Software 3.0:

• "LLMs + RL automate what is verifiable"
• Tasks with an automated reward signal = AI improves quickly
• math, coding, tests, benchmarks, games → areas of fast progress
• Conversely: tasks that are hard to verify remain a limit for AI automation

State of the Art (2026-07)

The paradigm stopped being an argument and became tooling. Three of the four Tier-1 labs now ship an agent framework as a product rather than a demo:

• Google ADK 2.0 + Agents CLI (2026-06-30) — the first model-agnostic end-to-end agent toolchain from a Tier-1 lab, explicitly built to run any framework
• Project Polaris and MAI-Code-1 / MAI-Code-1-Flash — Microsoft's agent platform and coding models, with Microsoft rebuilding its developer tools around the same premise
• OpenAI Parameter Golf — What It Taught Us (2026-05) is the quiet evidence: an OpenAI community ML challenge found coding agents had already become standard research tooling, lowering the barrier to ML research itself. The paradigm arrived in the research loop before anyone announced it

State of the Art (2026-05)

• Karpathy's framing: the most influential articulation of Software 3.0 in the industry
• Similar concepts: Agents (LLM Agents), Agentic Reinforcement Learning, Reasoning Models
• Distinction: a framework that describes a paradigm shift rather than a technology

Open Problems

• In Software 3.0, how to handle tasks outside the "verifiable" domain (creative work, strategy, judgment)?
• The quality-maintenance mechanism of agentic engineering needs standardization
• What is the competitive advantage of a company built on Software 3.0?

Key Sources

• Andrej Karpathy — Sequoia Ascent 2026 fireside chat (source)
• Karpathy blog: https://karpathy.bearblog.dev/sequoia-ascent-2026/
• X: https://x.com/karpathy/status/2049903821095354523

Related Concepts

• Agents (LLM Agents) — the execution unit of Software 3.0
• Agentic Reinforcement Learning — combining agentic engineering + RL
• Reasoning Models — beneficiary area of the verifiability principle
• LLM Knowledge Bases (LLM-curated personal wikis) — Software 3.0's knowledge management pattern (this system itself)