How Karpathy Envisions Software 3.0: Agents as the New Programming Paradigm

Andrej Karpathy, co‑founder of OpenAI and former Tesla Autopilot lead, now runs Eureka Labs focused on AI‑driven teaching. In a Sequoia Ascent fireside chat (April 30 2026) he explained how AI agents are transforming software and work.

He frames the transition as three software generations: Software 1.0 (human‑written code), Software 2.0 (models trained on data and objectives, with knowledge embedded in weights), and Software 3.0, where the LLM acts as an interpreter and the context window becomes the program. In this view, prompting, tools, examples and instructions replace explicit code.

Karpathy illustrates the shift with a menu‑generation (MenuGen) project. The traditional pipeline required a multi‑component stack—frontend, API, image generation, deployment, authentication, payment, key management. In the Software 3.0 approach a single textual prompt to an agent can produce the final image‑augmented menu, eliminating the conventional software scaffolding.

He introduces a "verifiability + training attention" framework: a model’s ability spikes in domains that are both verifiable (have clear reward signals) and heavily emphasized during training. This produces a "zig‑zag" intelligence curve—rapid progress in math, coding, testing, games, and stagnation elsewhere.

Karpathy proposes a rough formula where ability ≈ verifiability × training attention × data coverage × economic value, citing the dramatic chess skill jump from GPT‑3.5 to GPT‑4 as an example of added training data.

He distinguishes Vibe Coding (low‑bar, descriptive prototyping) from Agentic Engineering (high‑bar, professional coordination of agents). Agentic engineers design specifications, supervise plans, write tests, manage permissions, isolate work trees, and ensure quality, rather than blindly accepting generated code.

Through the MenuGen payment‑bug example, he shows why human judgment remains essential: an agent matched Stripe purchase records to Google accounts via email, ignoring the possibility of mismatched emails, highlighting the need for persistent user IDs.

Karpathy stresses that programmers are becoming agents’ orchestrators, not code writers, and that interview processes should evolve to evaluate a candidate’s ability to delegate tasks to agents, build specifications, and secure deployments.

He predicts a future where every person and organization has a personal agent that interacts with other agents to schedule meetings, execute tasks, and manage workflows, turning software, research, education, and knowledge work into variations of the same agent‑centric model.

For entrepreneurs, he advises seeking valuable, verifiable problems that large labs have not yet prioritized, creating environments where models can receive reliable reward signals, then fine‑tune or reinforce‑learn to achieve high performance.

Finally, Karpathy warns against anthropomorphizing LLMs; they are statistical artifacts shaped by pre‑training, fine‑tuning, reinforcement learning, product feedback and economic incentives, exhibiting impressive capabilities in some moments and glaring errors in others.

dim

axis

keepdim

reshape

permute