Agentic engineering and Karpathy's Software 3.0

Andrej Karpathy spoke with Sequoia partner Stephanie Zhan at AI Ascent 2026. He published a summary and an edited transcript on his blog on April 30. He notes that an LLM produced the summary from the transcript and that he reviewed it. The talk gives founders a working model for building software with agents.

December 2025 as an inflection point

Karpathy says coding agents became reliable around December 2025. Generated code arrived in larger, more coherent chunks, and he corrected it less. The unit of programming shifted from typing lines to delegating larger tasks: implement a feature, refactor a subsystem, write tests and fix failures. He describes the programmer as increasingly an orchestrator of agents.

Software 3.0

Karpathy frames three eras. In Software 1.0, humans write explicit code. In Software 2.0, humans train neural networks on data. In Software 3.0, humans program LLMs through prompts, context, tools and examples. The context window becomes the main lever.

His example is MenuGen, an app he built to show photos of dishes on a restaurant menu. It needed a frontend, image generation, payments and deployment. He later saw a multimodal model do the same job in one step from a photo. His module for founders: some apps should stop existing as apps.

Automate what you can verify

Karpathy's core rule: traditional software automates what you can specify, and LLMs automate what you can verify. Tasks with a clear success signal, such as code that passes tests, improve fastest because models can practise them. He adds a second factor. Capability also depends on how much attention labs gave a task in training. He calls the result jagged intelligence.

For founders, he points to domains that are valuable, verifiable and not yet a focus of frontier labs. A team that builds its own evaluation environment in such a domain can improve results with fine-tuning or reinforcement learning.

Vibe coding versus agentic engineering

Karpathy draws a clear line. Vibe coding raises the floor, letting almost anyone build software by describing it. Agentic engineering raises the ceiling. It is the discipline of coordinating fallible agents while keeping correctness, security and maintainability.

He gives a concrete failure. An agent matched Stripe purchases to user accounts by email address, but a customer's Stripe email and Google login can differ. A human needed to insist on persistent user IDs. Agents recall API details well. The engineer still has to understand the system.

Build for the agent, not only the human

Karpathy says most documentation still tells humans where to click. He wants instructions an agent can follow. He lists agent-native surfaces: Markdown docs, CLIs, APIs, MCP servers, structured logs and machine-readable schemas.

Hiring for agentic skill

Karpathy argues that puzzle-style coding interviews test the old paradigm. He suggests giving a candidate a substantial project to build and deploy with agents, then having other agents try to break it.

Founder takeaways

• Write specs before prompts. Have agents draft the spec, then review it yourself.
• Add tests to every agent task. Verification is what makes agent output trustworthy.
• Publish agent-ready docs. If your customers use agents, your product needs a CLI, an API or an MCP server.
• Change your technical interview. Test how candidates direct and review agents.
• Keep understanding in-house. Karpathy quotes a line he likes: you can outsource your thinking, but not your understanding.

Sources

1. Sequoia Ascent 2026 summary
   Andrej Karpathy (karpathy.bearblog.dev) · Apr 30, 2026
2. AI Ascent 2026
   Sequoia Capital · May 8, 2026