Agentic Engineering shifts the developer from typing code to delegating tasks. You give AI agents clear goals and constraints. They plan, execute, and report back while you supervise.
This page defines the term, traces its origin, and explains the core principles. You will also see how it differs from Vibe Coding.
| Aspect | Agentic Engineering | Vibe Coding |
|---|---|---|
| Interaction | Delegate tasks with constraints | Describe features in natural language |
| Developer role | Supervisor, reviewer | Director, idea describer |
| Oversight | Explicit review gates | Trust but verify ad‑hoc |
| Scope | Multi‑step professional tasks | Prototypes and weekend projects |
What is Agentic Engineering?
Agentic Engineering means delegating software tasks to AI agents. You define the goal and the rules. The agent plans its steps, picks tools, and produces output. You review and approve.
It moves beyond vibe coding. Vibe coding is about fast prototyping. Agentic engineering adds structure, safety, and professionalism. Andrej Karpathy framed it this way at Sequoia’s AI Ascent conference.
Learn who coined the term and the full origin story.
Who coined Agentic Engineering and when?
Andrej Karpathy introduced “programming via LLM agents” at Sequoia’s AI Ascent in March 2026. He described it as the next step after vibe coding. The community quickly adopted the name Agentic Engineering.
Karpathy’s talk outlined a vision where developers manage a team of AI agents. Each agent handles a specific part of the work. Humans stay in the loop for oversight and key decisions.
Read the detailed origin story from the AI Ascent keynote.
What are the core principles of Agentic Engineering?
Three principles anchor the discipline. First, delegation: give agents clear task contracts and let them work. Second, supervision: maintain human review gates for all critical outputs. Third, oversight: log, audit, and rollback when needed.
These principles separate Agentic Engineering from less structured AI coding. You don’t just trust. You verify. You don’t just prompt. You manage.
How is Agentic Engineering different from Vibe Coding?
Vibe Coding is about speed and exploration. You describe an idea. AI generates code. Agentic Engineering adds layers of control. You delegate a task with constraints and review gates. The agent plans, executes, and reports back.
Karpathy described Vibe Coding as a step toward Agentic Engineering. The former lets you play. The latter lets you build professionally.
See a full comparison between the two approaches.
What exactly is an AI agent in this context?
An AI agent is an LLM‑powered system that can set sub‑goals, call tools, and maintain memory. It works in loops. It plans, takes an action, observes the result, and adjusts.
Not every LLM call is an agent. A simple autocomplete is not. An agent autonomously executes multi‑step tasks. It evaluates progress. It retries on failure.
Understand the full definition of an AI agent here.
Where does context engineering fit into this?
Agents need rules. Context engineering provides those rules through files like AGENTS.md or CLAUDE.md. These files sit in the project root. Every agent reads them before acting.
Good context files specify the tech stack, coding standards, and forbidden actions. They turn a generic agent into a specialized, safe team member.
Learn how to build effective context files.
(1) Coined by Andrej Karpathy at AI Ascent 2026.
(2) Based on delegation, supervision, and oversight.
(3) Delegates tasks to AI agents that plan and execute.
(4) Different from Vibe Coding in structure and professionalism.
(5) Requires context files and guardrails to be safe.
References
- Andrej Karpathy, “Programming via LLM Agents” — AI Ascent, Sequoia Capital, March 2026
- Who Coined Agentic Engineering? The Full Origin Story
- Agentic Engineering vs Vibe Coding — Key Differences
- What Is an AI Agent? Goals, Tools, Memory & Autonomy
Always refer to the official documentation.
Leave a Reply