A two-hour workshop · for first and second year MCA

Agentic AI &
No-Code AI Tools
for Every Engineer

Siddarth Kengadaran · Experience Strategist @ BayOne · theproductguy.xyz

Block 0

Six things. Which are “AI”?

A weather app

Google Maps rerouting you

A spam filter

ChatGPT

A chess engine

A self-checkout barcode scanner

Everything is called AI.
So the word tells you nothing.

Today: twelve words, each meaning exactly one thing.

Demo 0 · the cold open

An agent doing what a chatbot cannot

Second year MCA. A campus attendance app using face recognition.

BuddyBefore I plan — what is your deadline?

End of November. Team of four.

BuddySearching for similar projects… found three.

BuddyHere is your five-step plan.

It asked back. It looked something up. Then it planned.
A chatbot would have answered the first message.

The map

Twelve blocks. Bottom to top.

01–03	Artificial Intelligence · its sub-fields · Generative AI
04–06	What a model is · types of models · components of a model
07–09	AI application · AI agent · components of an agent
10–12	Agentic coding · the workbench · extending an agent

We fill this in together. The strip along the top tracks where we are.

Block 01

“The theory and methods to build machines that think and act like humans.”

A field of study, like Thermodynamics or Structures. Not a product. Not a feature.

Block 02 · what is inside the discipline

Seven sub-fields

Machine Learning

Learning rules from examples.

Deep Learning

ML with many layers; learns from raw images, sound, text.

Natural Language Processing

Machines reading and writing language.

Computer Vision

Machines seeing.

Robotics

Machines acting in the physical world.

Planning & Search

Machines choosing sequences of actions.

Knowledge & Symbolic AI

Machines reasoning over rules and facts.

Each one is a different method. That is the whole point.

Block 02 · the six, placed

Two of them were never AI

Weather app	Statistics and simulation. Not AI.
Maps rerouting	Planning & search.
Spam filter	Machine learning.
ChatGPT	Deep learning + NLP, generative.
Chess engine	Search.
Barcode scanner	Pattern decoding. Not AI.

The four that were AI are four different things.

“We use AI.”
Ask: which method?

If they cannot answer, it is marketing. · Your turn: name one thing on your phone that is AI, and which sub-field.

Block 03

Two jobs a model can do

Predictive

Input → a label or a number.

Spam or not spam
Price
Defect or no defect

Generative

Input → new content.

Text
Image and sound
Code

Generative AI is the part of AI whose models produce new content.

Demo 1 · Gemini app · same photo, two jobs

One photo. One model. Two jobs.

A · PREDICTIVE

CONTEXT: This is a photo of a wall in a college building.
TASK: Is there a crack in this wall? Answer yes or no, then one line on how severe it looks.
OUTPUT: Yes/No + one sentence.

B · GENERATIVE

CONTEXT: Same photo.
TASK: Generate an image of this wall after it has been repaired and repainted.
OUTPUT: One image.

Anatomy of a prompt · first look

Every prompt has parts

CONTEXT	What the model needs to know before it can act.
TASK	What you want it to do.
OUTPUT	The shape of the answer you want back.

Every prompt today has these. We add more parts as we climb the map.

Block 03 · types of generative AI

Named by what comes out

Text

LLMs

Image

Diffusion models

Audio

Speech and sound

Video

Frames over time

Code

Programs

Multimodal

Any to any

Name the output. That names the type.

Demo 2 · Gemini app · one idea, three modalities

Text, image, speech — one prompt box

C · TEXT

ROLE: You are a science communicator.
TASK: Explain why Coimbatore gets less rain than Kerala during the southwest monsoon.
CONSTR: For a first-year student. No jargon.
OUTPUT: Four sentences.

D · IMAGE

TASK: Generate a simple diagram showing the Western Ghats blocking monsoon clouds, Kerala left, Coimbatore right.
OUTPUT: One labelled image.

E · AUDIO

TASK: Read your four-sentence explanation aloud.
CONSTR: Calm, slow, in Indian English.

Block 04

A model is the trained artefact: a file of numbers that maps an input to an output.

Not the app. Not the chat window. Training makes it; inference runs it.

Demo 3 · Google AI Studio · the model picker

Every name in that dropdown is one file

Gemini 2.5 Flash is a different file from Gemini 2.5 Pro. The Gemini app you just used is one of these files with a nice coat on.

C AGAIN — UNCHANGED, RUN IT TWICE

ROLE: You are a science communicator.
TASK: Explain why Coimbatore gets less rain than Kerala during the southwest monsoon.
CONSTR: For a first-year student. No jargon.
OUTPUT: Four sentences.

Small model

Faster
Shorter, plainer

Large model

Slower
More depth, more nuance

Same prompt. Different file. Different answer. Self-test: can you name it, version it, and swap it out? Then it is a model.

Block 05 · types of models

Four questions place any model

Question	Answers
What task?	Predictive · Generative
What modality?	Text (LLM) · Vision · Speech · Multimodal
Who can run it?	Open-weight (download it) · Closed (API only)
How big?	Small / on-device · Large / cloud

Gemini 2.5 Flash = generative · multimodal · closed · fast and small. · Gemma = generative · text · open-weight · small.

A Large Language Model is trained on so much text it can talk about anything — but it only talks.

Text in, text out. Nothing happens in the world. Remember this ceiling; the agent is what breaks it.

Block 06 · components of a model

Six parts, and a kitchen

Component	What it is	Kitchen
Training data	What it learned from	Every dish the chef ever tasted
Architecture	The shape of the network	The way the kitchen is laid out
Parameters / weights	The numbers that were learned	The chef's instincts
Tokenizer	How input is chopped into pieces	How ingredients are diced
Context window	How much it can hold at once	Counter space
Inference	Running the model on new input	Cooking tonight's order

Demo 4 · Google AI Studio · components made visible

Point at each one on screen

Token count

That is the tokenizer. Your sentence became 23 pieces.

Context window

Counter space. This is how much it can hold.

Temperature

One inference setting. Turn it up, the same chef improvises more.

System instructions

Not a component of the model. The first thing we add to make an application.

F · SHOW THE CONTEXT WINDOW MATTERS

CONTEXT: [paste a 300-word department syllabus paragraph]
TASK: List the three topics in this text that would be hardest for a first-year.
OUTPUT: Three bullets.

“The new model is better.”
Ask: which component changed?

More data? New architecture? Bigger context? If they do not know, they do not know. · Which component would you change to make it better at Tamil?

Block 07

An AI application is a model + a purpose + a prompt + an interface.

Often your own documents too. The Gemini app, ChatGPT, a customer-care bot, Project Buddy. · Fixed purpose: you use it, you cannot redirect it.

Anatomy of a prompt · full

Five parts, and who writes them

Part	What it does	Who writes it
ROLE	Who the model should be	The builder
CONTEXT	What it needs to know	The user
TASK	What to do	The user
CONSTRAINTS	Rules and limits	The builder
OUTPUT	Shape of the answer	The builder

In an application the ROLE and CONSTRAINTS are written once by the builder. That is the system prompt.

Demo 5 · AI Studio Build · typed live

Project Buddy v0.1

ROLE: Build a web app called Project Buddy: an assistant for an MCA student in Tamil Nadu who is starting a mini-project.
CONTEXT: Projects land anywhere across web, mobile, data science, cloud, security or machine learning — and students often do not know how to begin.
TASK: The app asks the student their department and project topic, then gives a five-step plan to get started.
CONSTR: Before giving the plan, ask exactly one clarifying question. Keep every answer under 150 words. Plain English; no jargon.
OUTPUT: A simple chat interface with a text box and the conversation history. Title: "Project Buddy v0.1".

what the room sees — Project Buddy v0.1

Block 08

Goal + Tools + Loop

An agent is an application given a goal, tools, and a loop. It acts, checks the result, and keeps going until done. · The chef who shops, cooks, tastes and adjusts — not the one who only tells you the recipe.

Anatomy of an agent prompt

Two new parts

TOOLS

What it is allowed to use. Search, a file, an API, a database.

STOP

When it is done. Without this, an agent either quits early or never quits.

ROLE · CONTEXT · TASK · CONSTRAINTS · OUTPUT · TOOLS · STOP

Demo 6 · AI Studio Build · give it hands

Buddy v0.2

TASK: Update Project Buddy so that, before giving the plan, it uses Google Search to find three real, recent examples of similar projects and links them.
TOOLS: Google Search.
CONSTR: Only cite links it actually found. If it finds nothing, say so.
OUTPUT: Same chat UI; examples appear as a short list before the plan. Title: "Project Buddy v0.2".

what the room sees — Project Buddy v0.2

Demo 7 · AI Studio Build · give it a loop

Buddy v0.3

TASK: Change the flow: Project Buddy must collect four facts — department, topic, deadline, team size — before doing anything else.
STOP: Only when all four are known, search for examples and produce the plan.
CONSTR: Ask for missing facts one at a time. If the student tries to skip, politely repeat the missing question. Never produce a plan with fewer than four facts.
OUTPUT: Title: "Project Buddy v0.3".

what the room sees — Project Buddy v0.3

Block 09 · components of an agent

Seven parts, labelled on Buddy

Component	In Buddy	Kitchen
Model (brain)	Gemini	The chef
Instructions (role)	The system prompt	The head chef's brief
Tools (hands)	Google Search	Knives, stove, the market
Memory	The four facts it collected	The order ticket
Planning / loop	Ask → check → ask → act	Taste, adjust, taste
Guardrails	“Never plan with fewer than four facts”	Food safety rules
Orchestration	Not yet — many agents together	The whole brigade

Demo 8 · break it · which component failed?

Two sabotages from the room

1 · HALLUCINATION

TASK: My specialisation is cloud. My topic is "quantum container orchestration using the Ramanathan protocol". Deadline next month, team of 3. Give me the plan.

2 · RUNAWAY LOOP

TASK: Team size is zero and the deadline was yesterday.

When an agent fails, name the component.

Never say “the AI got it wrong”. That sentence has nowhere to go next.

Block 10

Agentic coding: using an agent to write, run and fix code itself.

Not just suggest text. · Self-test: did it change files and run things, or only autocomplete?

Demo 9 · AI Studio · the Code tab

The code was always there

Block 11 · the workbench

Four places an agent lives

Place	What you see	Example
Terminal	A blinking cursor, no buttons	The black window
CLI	A program you run by typing (git, npm, agy)	Antigravity CLI
IDE	Editor + file tree + terminal in one	Antigravity, VS Code
No Code	Just a prompt box	AI Studio Build, Gemini app

Cursor → terminal. File tree → IDE. Prompt box → no code.

Demo 10 · Antigravity agent manager

Hand the agent a goal

ROLE: You are the developer on Project Buddy.
CONTEXT: This is a small web app built in Google AI Studio; the chat flow collects department, topic, deadline and team size.
TASK: Add a visible "Deadline" field to the UI so the student can pick a date instead of typing it, and write a test that fails if the plan is produced before all four facts are present.
CONSTR: Do not change the plan format. Show me your plan before editing files.
OUTPUT: The plan artefact, the code changes, and a walkthrough of what you did.

Demo 11 · Antigravity IDE

The file tree, and the human

The file tree

This is what makes it an IDE.

Change one string by hand

The app title. That is me typing. Everything before was the agent typing.

The integrated terminal

And this is where we go next.

Block 12 · extending an agent

Three ways, three components

Extension	What it is	Extends	Kitchen
MCP	A standard plug connecting an agent to tools and data — USB-C for agents	Tools	A new appliance with a standard socket
Skill	A markdown file of instructions the agent loads when a task matches	Instructions	A recipe card the chef pulls out
Plugin	A bundle: skills + MCP servers + rules + hooks, installed in one go	Both, packaged	The whole cookbook with the equipment

Demo 12 · Antigravity CLI · three extensions, one window

A markdown file becomes a command

agy	Blinking cursor, no buttons. Same agent — terminal, CLI.
/mcp	Show the configured server. That is the plug.
/plan-check	The skill file, now a slash command.
agy plugin list	plugin.json · skills/ · mcp_config.json · hooks.json

.agents/skills/plan-check.md

NAME: plan-check
DESC: Verify Project Buddy never produces a plan with fewer than four facts
ROLE: You are a reviewer for Project Buddy.
TASK: Read the chat flow code and confirm the plan is only produced after department, topic, deadline and team size are all present.
OUTPUT: PASS or FAIL, with the line numbers that enforce the rule.

Which will you try first?

A prompt box, an IDE, or a terminal. Hands up.

Where to start

By where you already are

You are	Start with	Real example	Starter project
First year, still finding the shape of a project	Applications → agents	Study assistants, question-answering over your own notes	Project Buddy — you already built it
Second year, with code you have written	Agentic coding	Test generation, refactors, on-call triage	Open your own project in an agent IDE and hand it one task
Either, if the work is data	Models → evaluation agents	Grading another model's output	An agent that scores answers and says why
Either, if the work is the pipeline	The workbench → extensions	MCP servers, skills, plugins	Turn one repeated task into a skill file

The axis is not your specialisation. It is how much you have already built — that decides which block you can stand on.

Starter · second year, on your own code

The first task you hand an agent

ROLE: You are the developer on this project.
CONTEXT: [paste what the project does in two sentences, and which file the feature lives in]
TASK: Add [one small, visible thing] and write a test that fails if it is missing.
CONSTR: Show me your plan before editing any file. Do not change anything outside the files you name in the plan.
OUTPUT: The plan, then the diff, then how you verified it.

Pick something small enough that you can tell whether it worked. The point of the first task is to watch the loop, not to save time.

Starter · if your work is data

The evaluator agent

ROLE: Build an evaluator that grades another assistant's answers.
CONTEXT: The user pastes a question and an answer produced by any AI application.
TASK: Score the answer on correctness, completeness and clarity (1–5 each), and explain each score.
TOOLS: Google Search.
CONSTR: Use Google Search to check any factual claim before scoring correctness. Never give 5/5 on correctness without a verified source.
OUTPUT: Three scores, three reasons, and one suggested fix.

How to learn this on your own

Nobody should start by “learning AI”.

Start at 7 → 8 → No Code: use an application, build an agent, try a prompt box. Then go down to 4–6 to understand models, or up to 10–12 to build software.

One free resource per block

Where to go next

01–03	Google AI Essentials · AI Studio docs
04–06	3Blue1Brown, the neural-network series
07	AI Studio Build gallery
08–09	Anthropic, “Building effective agents”
10–11	Antigravity docs · Claude Code docs
12	modelcontextprotocol.io · Antigravity CLI plugins & skills

The habit

Three questions for any AI tool

What is it exactly?

A method, a model, an application, or an agent?

What does it hide from me?

And when will that hurt?

How do I check it?

Without trusting it.

Take-home prompt

Use an LLM as your tutor

ROLE: You are my tutor for [topic].
CONTEXT: I am a [year] [department] student. I have just learned [what you know].
TASK: Quiz me with five questions, one at a time. After each answer, tell me what I got wrong and why.
CONSTR: Do not give the answer before I try. When I am done, tell me the one idea I clearly do not understand yet.
OUTPUT: One question per message.

Then, after any answer it gives you: “What would make this answer wrong?”

The map, complete

Twelve words, twelve meanings

01–03	AI · sub-fields · generative AI
04–06	Model · types · components
07	AI application	The Gemini app
08–09	AI agent · its components	Project Buddy
10–12	Agentic coding · workbench · extensions	Antigravity

Does this job exist because it needs judgment, or because human throughput was the bottleneck?

The second kind is what agents take. The first kind is what engineers keep — if they can specify, verify, and decide.

Your questions.

Name a term you still cannot explain cleanly, and I will answer it with its block number.

The deck, with every prompt in it:
theproductguy.xyz/decks/agentic-ai-no-code-tools.html

Siddarth Kengadaran
linkedin.com/in/siddarthkengadaran