Skip to content

Agentic Engineering

Coined by andrej-karpathy at Sequoia AI Ascent 2026, as the professional counterpart to vibe-coding:

"Agentic engineering is about preserving the quality bar of what existed before in professional software. You're not allowed to introduce vulnerabilities due to vibe coding. You are still responsible for your software just as before, but can you go faster?"

Definition

The engineering discipline of coordinating LLM agents — "spiky entities, a bit fallible, a little bit stochastic, but extremely powerful" — to ship faster without sacrificing the quality bar.

What it actually looks like

  • Tooling investment. Like vim/VS Code before it: mastery = getting the most out of Claude Code / Codex / Cursor / similar, plus your own harness.
  • Staying in the loop. Treat agents as tools, not replacements. You own taste, architecture, and design; the agent owns fill-in-the-blanks.
  • Review discipline. Agents mis-correlate things humans wouldn't (Karpathy's Stripe↔Google email-matching bug) — so review surface grows, doesn't shrink.
  • Agent-native infrastructure. Docs, APIs, data structures should be legible to LLMs first; humans second. Current world is backwards ("why are people still telling me what to do? What's the thing I should paste to my agent?").

Ceiling claim

"People used to talk about the 10x engineer. This is magnified a lot more. 10x is not the speed-up you gain."

Karpathy's strong prior: the ceiling of what a great agentic engineer can do is much higher than 10x, while the floor (vibe-coding) rises independently.