Previously, 
on Fullstack Carrot…

Quantum recap: Not Either / Or

The future is probably
not classical or quantum.
It is classical and quantum.

I’m not a faster computer.
I’m a different
kind of computer.

WAIT! Before the story begins…
What’s your AI level?

Find yourself on the AI Mastery 
chart below. 
Then let’s see where this 
story takes you!

Bonus

A Bakery Has a Problem

Monday morning. 
One small bakery.
One big weekend
campaign

Can AI help me plan the 
campaign, check inventory, 
draft posts, and build
a tiny landing page?

A chatbot can answer. 
An agentic system can 
help move the work
forward

First: What a Model Does

A generative model
turns patterns from
training into a likely
next piece of text,
code, image, or sound.

It is brilliant at
generating possibilities.
But by itself,
it cannot reliably 
know what is true
—or take actions

That gap is where
agent design begins

Foundation Models: The Engine

Foundation models
are large, broadly
trained engines.

Some are
optimized
for language.
Some also
see images,
hear audio,
write code,
or handle
long context.

Choose a model
for the job
—not for a mascot,
a benchmark
screenshot,
or a viral post.

Bonus

Inside the Language Machine

The model does not
read words exactly
as we do.
It breaks input
into token-sized pieces.

Those pieces become
vectors: 
locations in a giant
mathematical
landscape.

Then it predicts
a plausible
next token
—again and again.

Training teaches
the model patterns.
At runtime, an agent
supplies goals,
context, tools, and rules.

Bonus

These are cousins 
of LLM, specialists.

The Agent Toolkit

An agent is not
one magic prompt.
It is a system

For the bakery mission,
Carrot needs a goal, trusted
context, tools, a working
memory, and boundaries.

The most important
component is
often the least
glamorous:
permissions.

FYI, AI job seeks!

Bonus

RAG: Ask Before You Answer

The bakery’s facts live
outside the model:
inventory, hours, recipes, 
campaign notes, and policies.

Classic RAG
retrieves
relevant 
chunks,
then gives
them to the
model
as context.

Agentic RAG decides
what to look up next, 
checks the evidence,
and can try again

Retrieval helps.
It does not
automatically make
an answer true.

Vectors and embeddings
are the superstars
behind RAG!

Bonus

AI technology is 
advancing incredibly fast.

The Agent Loop

Now the system can loop:
understand the goal, plan,
act, observe, and decide
what comes next

Goal: 
“Prepare the bakery’s
weekend campaign.”

Plan: 
break one fuzzy request
into smaller, checkable tasks.

Act: 
call a permitted tool. 
Observe: inspect 
what happened.

Repeat—or ask a human
when the task needs
judgment, permission,
or accountability.

Bonus

Shampoo,
rinse
and repeat.

Memory Without Magic

An agent needs
 enough short-term
memory to stay on task

Long-term memory
is optional—and should 
be deliberate, useful,
and permissioned

Good memory says:
“What should we retain?”
Not: “Let’s keep
everything forever.

Bonus

This is not just theory. Recently, long
conversations and context-management
problems caused some very
public trouble for Google’s Gemini

One Agent or Many?

Start with one capable
agent and clear tools

Add specialists only 
when the work truly
benefits from delegation.

More agents can mean more capability
—and more coordination,
cost, failure modes, and debugging.

Bonus

Choose Multi-agent when the task 
needs genuinely separate skills,
tools, or parallel work
—not just because more agents
sound smarter.

Guardrails Are the Product

A useful agent must be 
able to stop, explain 
what it did, and 
how its evidence

Never give broad, irreversible 
power to a system that has
not earned your trust

Evaluate the workflow, 
not just the model: quality, 
cost, latency, reliability,
and safety

The Result: Assisted, Not Replaced

The agent gathers facts,
drafts a campaign, checks
stock, and prepares 
a simple launch plan.

The owner reviews the result,
changes the creative direction,
and gives final approval.

Agentic AI is not “AI replaces the human.”
It is “AI helps the human direct more work.

Bonus: “Vibe Coding, Carefully”

Bonus round: 
“Can we build the bakery page
by describing the vibe?”

Vibe coding: you express
intent in natural language,
and AI helps generate
and revise code.

Great for prototypes. 
Still test, review, secure,
and understand
what you ship.

Next episode: 
Fullstack Carrot goes vibe coding
—without losing the plot.