What is agentic engineering?
Agentic engineering is the discipline of directing fallible coding agents without lowering the quality bar, a name Andrej Karpathy popularised on 4 February 2026. Instead of asking whether an agent can write something, you place each loop on the autonomy ladder and ask what would move it up, such as a stop condition a machine can check instead of a human reading the diff.
Origin Andrej Karpathy popularised the name on 4 February 2026 in a post on X (opens in a new tab) that called “agentic engineering” his current favourite name for the practice, after the phrase was already in use, as in Mario Zechner’s post of 2 June 2025 (opens in a new tab).
- In Polish
- Inżynieria agentowa
- Closest ladder level
- All levels, free guide
$59.99 once for 12 months of access. Nothing renews on its own. 30-day refund on a first purchase.
Why it matters.
The name changes the question. "Can an agent write this?" has no useful answer; "which level is this loop at, and what would move it up?" does. Each level describes one loop, so a team has no single level, and from Level 3 upward a loop climbs only when a machine-checkable signal, such as commands that exit 0, replaces a human reading the diff.
Agentic engineering vs vibe coding
Both put agents in charge of the typing. They differ in who answers for the result and what ends the loop.
| Criterion | Vibe coding | Agentic engineering |
|---|---|---|
| What it names | A practice: prompt, glance at the result, accept it | A discipline: directing coding agents without lowering the quality bar |
| Who writes the code | The model | Agents, which the human orchestrates and oversees |
| Who checks it | Nobody reads the diff; the human watches behaviour | The human stays accountable; tests and evals check where they can |
| What stops the loop | “It mostly works” | The quality bar is met |
| Where the name came from | Andrej Karpathy, X post, 2 Feb 2025 | Earliest use found: 2 Jun 2025 (Mario Zechner); Karpathy popularised it on 4 Feb 2026 |
| Where it breaks | When the software is for other people: Willison calls it “grossly irresponsible” there (6 May 2026) | When the boundary erodes: Willison finds the two have “started to blur” and sees “normalization of deviance” (6 May 2026) |
The six levels of the autonomy ladder
Each level describes one loop, a repeatable class of change with its own trigger, check and stop condition, not a whole team.
| Level | Who writes | Who reads | In Shapiro’s words |
|---|---|---|---|
| L0 · By hand (Spicy autocomplete) | Human | Human, every character | "not a character hits the disk without your approval" |
| L1 · Assisted (The coding intern) | Human, mostly | Human, all of it | "you offload specific, discrete tasks to your AI intern" |
| L2 · Paired (The junior developer) | Human and agent, paired | Human, every line | "feels like you are done. But you are not done" |
| L3 · Review manager (The developer) | Agent, most of it | Human, as full-time reviewer | "Your life is diffs." |
| L4 · Spec manager (The engineering team) | Agent, from a spec | Tests, mostly | "leave for 12 hours, and check to see if the tests pass" |
| L5 · Dark factory (The dark software factory) | Agent | Nobody | "It's a black box that turns specs into software." |
How each tool does it.
Cursor, Claude Code and Codex each name this their own way. Each note carries the date it was checked.
Cursor
/goal is marked "Rolling out" and /loop is a bundled skill since 3.5; hooks exchange JSON over stdio before or after agent-loop stages and can block.
Cursor docs, 2026-08-28
Claude Code
Keeps a run going with /goal, judged by a separate evaluator model each turn, and with the bundled /loop skill; hooks can block an action, with 33 events documented.
Feature matrix, Claude Code 2.1.283, 2026-09-26
Codex
Keeps a run going with /goal (feature goals, stable); hooks (12 events) need persisted trust before they run, and there is no /loop in the 0.157.1 binary or in 0.162.1 (checked 2026-10-12).
Feature matrix, Codex CLI 0.157.1, 2026-09-26
Questions about agentic engineering.
How is agentic engineering different from vibe coding?
In the transcript of his Sequoia Ascent talk that Karpathy published, he sets them against each other: "Vibe coding raises the floor. Agentic engineering is about extrapolating the ceiling." Vibe coding is accepting agent output without verifying it, which is fine for a throwaway prototype; agentic engineering keeps the quality bar while the agents do the typing.
Who coined agentic engineering?
No single person: Andrej Karpathy popularised the name on 4 February 2026, calling "agentic engineering" his current favourite, but Mario Zechner was already describing what people call agentic engineering on 2 June 2025 and Peter Steinberger titled a post with it on 14 October 2025. In the transcript of his Sequoia Ascent talk that Karpathy published on 30 April 2026, he calls agents "spiky entities" that are "fallible and stochastic, but extremely powerful", and asks how to coordinate them without sacrificing your quality bar.
Does a team have one level?
No. A level belongs to a loop, a repeatable class of change with its own trigger, check and stop condition. A dependency-bump loop can run at Level 4 while feature work in the same repository sits at Level 2.
Do I need a particular tool to practise it?
No. The feature matrix lists subagents, worktrees, skills, hooks, MCP, a headless mode, an SDK and scheduled runs (cloud Routines in Claude Code, cloud Automations in Cursor, local desktop-app automations in Codex) for all three (Claude Code and Codex checked on 2026-09-26, Cursor from docs read on 2026-08-28). They differ in where the agent runs, which controls block an action rather than steer it, and what checks a result before merge, so test the rows that decide your choice on your own repository.
The vocabulary of AI-driven development
Every name below is defined against vibe coding: who writes the code, who checks it, and what stops the loop.
Read the long-form guide: agentic engineering vs vibe coding
Sources.
The primary sources outside this site that this page relies on.
- Post on X about the name “agentic engineering” (opens in a new tab) Andrej Karpathy, X
- Prompts are code, .json/.md files are state (opens in a new tab) Mario Zechner
- Just Talk To It - the no-bs Way of Agentic Engineering (opens in a new tab) Peter Steinberger
- Sequoia Ascent 2026 summary (opens in a new tab) Andrej Karpathy
Keep reading.
The guides that go deeper, and the terms and comparisons next to this one.
In the docs
Related terms
Does this page answer your questions?
One tap, no sign-in.
What are you here for?
What is holding you back?
How does it read?
Thanks. If you left an email, Michał will reply.
Answering your questions?
One tap each. Close whenever you like.
no sign-in
Read the guides in the same words.
Open every guide with the Toolkit. Each term is defined once here, and the guides use it the same way.
$59.99 once for 12 months of access. Nothing renews on its own. 30-day refund on a first purchase.
Michał replies, not a bot. The address is used for this reply only.