[2026] What Is Agentic Engineering? Coined by Karpathy, Google Agents CLI Transforms Production-Grade Development

Q: What is ADK?

**ADK (Agent Development Kit)** is Google's framework for agent development. Agents CLI teaches your coding agent how to write, evaluate, and deploy using ADK patterns.

"I want to build AI agents. But making ones that actually work is way harder than I imagined..."

Sound familiar? In 2026, one of the most talked-about terms in the AI industry is "Agentic Engineering." The person who coined it? None other than Andrej Karpathy — the legendary engineer who made his mark at Tesla and OpenAI.

A viral post on X (42.7K views, by @akshay_pachaar) highlighted an article titled "Karpathy's Agentic Engineering Finally Has Proper Tooling," reporting that Google has released a tool called "Agents CLI," suddenly making this new development methodology practical and accessible.

In this article, we'll break it all down in plain language — no jargon overwhelm — so even beginners can follow along.

Here's what you'll learn:

What "Agentic Engineering" actually means
How it differs from just "having AI build stuff for you"
How Google Agents CLI makes development dramatically easier
The best learning path for beginners

All of it, explained clearly.

What Is "Agentic Engineering"? Starting with the Basics

In a nutshell: it's a "systematic discipline for designing and developing AI agents at a production-grade level" — agents that are truly ready for real-world use.

Karpathy defined the term at Sequoia Capital's "Sequoia Ascent 2026" event. The critical point: this is distinctly different from "vibe coding" (coding by feel).

Karpathy's Sequoia Ascent 2026 talk summary

In Karpathy's own framing, Agentic Engineering is the discipline that separates production-grade agent development from casual vibe coding.

Aspect	Vibe Coding	Agentic Engineering
Mindset	"Just have the AI build something that kinda works"	Design specs, evaluations, and safety from the start
Evaluation	Glance at it — "looks good to me"	Automated evals with numerical scoring
Safety	Largely overlooked	Monitoring and verification are mandatory
Output	Prototype-level	Production-grade

The 3 Core Skills of Agentic Engineering

Karpathy identified three essential skills for production-grade agent development:

1. Spec Design

Articulating what you want the agent to do as a clear, unambiguous specification — not a vague wish.

Bad example: "Handle customer inquiries nicely" Good example: "For each inquiry: ① reference the refund policy ② guide through the applicable procedure ③ escalate to a human if unable to resolve. Never output personal information."

2. Eval Loops

Running an automated, repeatable checking mechanism that continuously verifies whether your agent is actually performing correctly. Don't settle for "seems to work fine" — measure it with scores, persistently.

3. Security Oversight

Monitoring and verifying that your agent isn't taking unexpected actions. For instance, preventing behaviors like "auto-linking a Stripe payment email to a Google account by matching email addresses" — actions that sound plausible on the surface but are wrong by design.

The Old Problem: "Tools Were Too Fragmented"

All of this sounds great in theory, but in practice it was incredibly tedious.

Building a production-grade agent meant juggling a chain of disconnected tasks:

Writing code in an editor
Scaffolding the project in a terminal
Testing behavior in a browser
Deploying via a cloud console
Setting up evaluations in a separate framework

In other words, you had to shuttle between 5+ different screens, stitching everything together yourself. That made the barrier to doing Agentic Engineering seriously far too high.

How Google Agents CLI Changed the Game

Enter Google's "Agents CLI."

Google Agents CLI GitHub repository

Agents CLI is a tool that injects "7 skills" into your coding agent — whether that's Claude Code, Cursor, Codex, Antigravity, or others.

These 7 skills bundle the following knowledge:

ADK code patterns (how to write with Google's Agent Development Kit)
Project scaffolding
Evaluation setup (LLM-as-judge scoring)
Deployment configuration (for Agent Runtime or Cloud Run)
Observability (tracing agent behavior with Cloud Trace)

Google Developers Blog on Agents CLI

Why It's a Big Deal

A single setup command installs all the same ADK expertise across every coding agent you use, in one shot.

What that means:

You never leave your editor
You just give instructions in natural language
Build → Evaluate → Deploy all run automatically

The "fragmented tools" problem collapses into one unified experience.

A Step-by-Step Walkthrough

The article demonstrates the full cycle by building a RAG agent (an AI that answers questions from internal knowledge) from scratch and deploying it:

Step 1: Install Agents CLI

uvx google-agents-cli setup

That's it — 7 skills injected into your coding agent.

Step 2: Build the RAG Agent

Tell your coding agent something like:

Build a RAG agent that ingests documents, retrieves relevant
context, and answers questions with source citations.
Use the ADK agentic_rag template with Gemini 3.5 Flash.

Since the agent already knows ADK patterns, it auto-generates the project — including citation support and Vector Search configuration, right out of the box.

Step 3: Test Locally

Launch ADK Web UI on localhost

A chat interface opens in your browser. You can test with real questions and verify: "Is it pulling the right context? Is it showing source citations like [source: 1003]?"

Step 4: Evaluation (Eval)

Use LLM-as-judge to auto-score answer correctness. No need to manually inspect every response.

Step 5: Deploy and Observe

Deploy to Cloud Run or Agent Runtime. Cloud Trace visualizes the agent's reasoning path — every step it took.

Google ADK (Agent Development Kit) official documentation

Recommended Learning Path for Beginners

Agentic Engineering is a broad field. Here's the suggested order for newcomers:

Route A: Grasp the Big Picture (Concept-First)

Read this article to understand what "Agentic Engineering" means
Read the Sequoia Ascent 2026 talk summary by Karpathy
Skim through the official Google ADK documentation

Route B: Get Your Hands Dirty (Implementation-First)

Install Google Agents CLI with uvx google-agents-cli setup
Pair it with Claude Code or Cursor and build a simple RAG agent
Run evaluations and get a feel for what the scores mean

Route C: Go Deep (Theory & Operations)

Study eval loop design
Learn observability and security monitoring techniques
Deploy to production in a real work or personal project

FAQ

Q1. Can non-programmers learn Agentic Engineering?

The core concepts (spec design, eval, security) are understandable even for non-engineers. That said, to actually build with Agents CLI, some familiarity with terminal commands and Python will make things much smoother.

Q2. Is Google Agents CLI free?

The CLI setup itself is free. However, deploying to Google Cloud (Cloud Run / Agent Runtime) may incur cloud usage fees. Local development is free to try.

Q3. Can I use it without Claude Code or Cursor?

Agents CLI works by injecting skills into a coding agent, so you'll need one of the supported agents (Claude Code, Cursor, Codex, Antigravity, etc.).

Q4. What is ADK?

ADK (Agent Development Kit) is Google's framework for agent development. Agents CLI teaches your coding agent how to write, evaluate, and deploy using ADK patterns.

Q5. Is vibe coding dead?

Not "dead" — for personal experiments and prototypes, vibe coding works fine. But when you're building an agent that will be used in production or by other people, the Agentic Engineering mindset (evaluation, safety) becomes essential.

Q6. Is the information in this article based on 2026 developments?

Yes. This article is based on the June–July 2026 release of Google Agents CLI and Karpathy's Sequoia Ascent 2026 talk. For the latest tool details, always check the official documentation.

Summary

Agentic Engineering is a methodology for building AI agents at production-grade quality. Coined by Andrej Karpathy, it centers on three core skills: spec design, eval loops, and security oversight.

Until now, fragmented tools made it cumbersome — but with the arrival of Google Agents CLI, you can now go from build → evaluate → deploy, all within a single coding agent.

First, grasp the concepts through this article
Next, install Agents CLI and play around
Then, experience firsthand why eval loops matter

Follow this sequence, and within weeks you'll be building production-grade agents with your own hands.

This article is for informational purposes only. Details about Google Agents CLI and ADK are based on official announcements and documentation. For the latest specifications, please refer to each official site.