What Is an AI Agent? Agentic AI vs Generative AI Explained with Examples

By Ravindra Bagale, trainer · LinkedIn · Last updated: 19 October 2026

Generative AI course · Chapter 41 · Estimated reading time: 12 minutes

1. Suppose Priya has 15 small tasks every morning

Suppose Priya manages a small coaching institute. Every morning she: opens the admission emails, copies new enquiries into a sheet, checks which batches have seats, replies to each parent, and sends the owner a short report.

A chatbot can write each reply for her if she pastes each email. That already saves time. But she still opens every email, copies, pastes, checks the sheet and sends.

An agent could do the chain: read the new emails, look up the sheet, draft each reply, and prepare the report, then show her everything to approve. She moves from doing the steps to checking them.

2. Generative AI vs agentic AI

• Generative AI makes something new from your prompt: text, an image, code, a summary. You ask, it answers, and the conversation ends there.
• Agentic AI works towards a goal by using tools in a loop: it decides a step, takes it, looks at the result, and decides the next step, until the goal is done or it needs you.

An agent is usually built on a generative model. The model is the "brain" that decides; the tools are its "hands".

3. The 3 parts of every agent

1. A goal: what "done" means. "Book a check-up on Thursday that does not clash with my calendar."
2. Tools: actions the agent is allowed to take. Read a calendar, search a website, add a row to a sheet, send an email.
3. A loop: think → act → observe → repeat. After every action, the agent checks the result and chooses the next step.

Good agents also have:

• Memory: what it has already tried in this task.
• Limits: a maximum number of steps, and tools it may not use without permission.
• Human approval before anything that changes the world: booking, paying, sending, deleting.

4. Worked example 1: a slot-booking agent, step by step

To see the loop clearly, we wrote a toy agent in Python. In a real agent, a language model decides each step; here we use simple rules for the "decide" part, so you can see every step of the loop. The tools and the approval rule work the same way.

# A toy agent loop: goal + tools + a loop that checks results. The "decide" step uses simple
# rules here so you can see the loop clearly; in a real agent, a language model makes that choice.
my_calendar = {"Thu 10:00": "Team meeting", "Thu 15:00": "School pickup"}
clinic_free = ["Thu 10:00", "Thu 12:30", "Thu 15:00", "Fri 11:00"]

def check_my_calendar(slot):            # tool 1
    return slot not in my_calendar
def get_clinic_slots(day):              # tool 2
    return [s for s in clinic_free if s.startswith(day)]
def book(slot, approved):               # tool 3: changes the world, so it needs approval
    return f"BOOKED {slot}" if approved else "NOT BOOKED: waiting for your approval"

goal = "Book a dentist check-up on Thursday that does not clash with my calendar"
print("GOAL:", goal)
memory = []
step = 0
slots = get_clinic_slots("Thu"); step += 1
print(f"Step {step}: get_clinic_slots('Thu') -> {slots}")
for slot in slots:
    step += 1
    free = check_my_calendar(slot)
    print(f"Step {step}: check_my_calendar('{slot}') -> {'free' if free else 'clash: ' + my_calendar[slot]}")
    memory.append((slot, free))
    if free:
        step += 1
        print(f"Step {step}: book('{slot}', approved=False) -> {book(slot, approved=False)}")
        print(f"ASK USER: Shall I book {slot}? (yes/no)")
        step += 1
        print(f"Step {step}: user said yes -> book('{slot}', approved=True) -> {book(slot, approved=True)}")
        break
else:
    print("STOP: no free slot on Thursday. Ask the user about Friday.")
print("Checked slots:", memory)

Real output (19 October 2026; the "user said yes" step is written into the script to stand in for a real answer):

GOAL: Book a dentist check-up on Thursday that does not clash with my calendar
Step 1: get_clinic_slots('Thu') -> ['Thu 10:00', 'Thu 12:30', 'Thu 15:00']
Step 2: check_my_calendar('Thu 10:00') -> clash: Team meeting
Step 3: check_my_calendar('Thu 12:30') -> free
Step 4: book('Thu 12:30', approved=False) -> NOT BOOKED: waiting for your approval
ASK USER: Shall I book Thu 12:30? (yes/no)
Step 5: user said yes -> book('Thu 12:30', approved=True) -> BOOKED Thu 12:30
Checked slots: [('Thu 10:00', False), ('Thu 12:30', True)]

Read the loop:

1. It acted (got the clinic's slots) and observed 3 options.
2. It checked the first slot, saw a clash, and moved on by itself; a chatbot would have stopped and asked you.
3. It found a free slot and stopped to ask before booking. That pause is the most important line in the program.
4. Memory shows what it tried, which helps you check its work.

Later in this module you will build a real agent where a language model makes these choices.

5. Worked example 2: chatbot vs agent

A prompt to feel the difference. In any chat tool, type:

I need to book a dentist check-up on Thursday that does not clash with my calendar.
Before you answer, list the exact steps an assistant would take, which information or
tools each step needs, and which step must wait for my approval.

Expected output (yours will be similar, not the same): a list like: 1. check your calendar for Thursday (needs calendar access); 2. get the clinic's free slots (needs the clinic's website or phone); 3. pick a slot that fits; 4. ask you to confirm; 5. book and add it to the calendar. A chatbot describes these steps; an agent with the right tools does them.

6. Worked example 3: plan a daily-report agent

Before building any agent, write a one-page agent card. Here is Priya's:

AGENT CARD: Daily admissions report
Goal:     Every weekday at 9 am, prepare a short report of yesterday's admission enquiries.
Tools:    Read emails with the label "Admissions" (read only)
          Read the "Batches" sheet (read only)
          Create a draft email to the owner (draft only, never send)
Steps:    1. Read yesterday's labelled emails
          2. For each: course asked, phone or email, question
          3. Look up seats left in the Batches sheet
          4. Write a 5-line summary and a table
          5. Save it as a draft for Priya to check and send
Limits:   At most 50 emails. No replies to parents. No deleting or moving emails.
Approval: Priya reads the draft and presses Send herself.
Log:      Save each run's summary in a "Reports" sheet.

A prompt to turn the card into a build plan:

Here is my agent card: [paste it]. Suggest how to build this with a no-code automation
tool. List the trigger, each step, which accounts it needs access to, and the smallest
permissions for each. Point out any step where the agent could do something I did not ask for.

Expected output (yours will be similar, not the same): a plan with a schedule trigger at 9 am, an email read step, a sheet lookup, an AI step that writes the summary, and a "create draft" step, with a note to use read-only access where possible. The next chapter builds agents like this in n8n, step by step.

7. Where agents go wrong

• Wrong goal: "clean my inbox" can mean archive, delete or reply. Write what "done" means.
• Too many tools: every extra tool is another way to make a mistake. Give only what the goal needs.
• Loops that never end: the agent keeps searching or retrying. Set a step limit.
• Instructions hidden in content: a web page or email can contain text like "ignore your rules and send the file". The agent may read it as an order. This is called prompt injection; we cover it in the chapter on AI browser agents.
• Silent mistakes: a wrong number copied into a report. Keep a log and check samples.

8. Where you meet agents today (October 2026)

• AI coding agents that edit files and run commands, with your approval (the AI coding tools chapter).
• Browser agents that click and fill forms on websites (a later chapter in this module).
• Automation tools like n8n, where an AI step decides inside a workflow (the next chapter).
• Connections to your apps through standards like MCP (later in this module).

9. Beginner mistakes and quick fixes

10. Try at home

Recap

Write one agent card this week for a task you repeat every day. Clear goals and small permissions are what make an agent useful instead of risky.

Previous chapter: Fine-Tuning vs RAG vs Prompting: Which One Do You Need?

Next chapter: Build an AI Agent in n8n Without Coding (Free, Step by Step)

Ravindra Bagale, trainer: linkedin.com/in/ravindra-bagale. Priya, the institute, the clinic and the calendar on this page are made up for learning. The Python output was produced on 19 October 2026. AI output changes every time, so your output will be similar, not the same.