Site navigation and theme and language settings

Definition

Agentic engineering

Agentic engineering is Andrej Karpathy’s name (April 2026) for the engineering discipline of working with coding agents that are "fallible and stochastic, but extremely powerful": coordinating them to go faster without sacrificing the quality bar. It replaces "can the agent write this?" with "which level do I run it at, and what harness, loop and verification does that level need?"

Last updated

Why it matters

The name matters because it changes the question. "Can an agent write this?" has no useful answer. "Which level am I running this at, and what does that level require?" has one you can act on: a project file the agent reads before it writes, a way to run without you watching, and checks you trust more than the model.

How each tool handles it

Cursor
Rules in .cursor/rules are the harness; the bundled /loop skill, added in Cursor 3.5, runs unattended, and /goal shipped on 19 August 2026, still marked "Rolling out".
Claude Code
CLAUDE.md is the harness and hooks are the gates; /goal and the bundled /loop skill both keep a session working turn after turn.
Codex
AGENTS.md is the harness; /goal is a documented developer command (checked 2026-08-28), and sandbox policy is set separately from approvals. There is no /loop.

Frequently asked questions

How is agentic engineering different from vibe coding?

Karpathy sets them against each other: "Vibe coding raises the floor. Agentic engineering is about extrapolating the ceiling." Vibe coding is accepting output you did not read; agentic engineering keeps the quality bar while the agents do the typing.

Who named agentic engineering?

Andrej Karpathy, in his Sequoia Ascent summary of 30 April 2026. He describes agents as "fallible and stochastic, but extremely powerful" and puts the discipline in one question: how do you coordinate them to go faster without sacrificing your quality bar?

Do I need a particular tool to practise it?

No. Cursor, Claude Code and Codex all ship the same three pieces in different shapes: a project file the agent reads before it writes, a way to run without a person watching, and a sandbox or approval policy that bounds what it may touch. What changes is the level you run at.