The Math Behind the Machine/ Unit 22 · How an Agent Decides Checks 0/0 Sign in
Unit 22 of 22 · by Prof. SaurabhFree preview

How an Agent Decides

Ask a chatbot for a train ticket and it tells you about trains. Ask an agent and it goes and gets one: it searches, reads, finds a waitlist, changes its mind, asks you before paying, and stops when the ticket is in your hand. This unit is about the loop that makes that possible, and about the simple maths of why it so often goes wrong.

In the full unit
≈ 60 min read + play 9 interactive widgets · a loop you can watch, a playground where you build the agent 7 inline checks 🧾 2 proofs, folded away ✍ 8 solved practice problems

Section 1 is free. Unlock the other 8 sections and the practice arena for ₹299 — or all nine paid units for ₹999.

Unlock Unit 22 · ₹299Pricing

← Unit 21 · Build a GPT

1

A chatbot answers. An agent acts.

Imagine this

It is Wednesday night. You need to travel to Dehradun on Friday. You open an AI app and type one line:

"Get me a train ticket from Delhi to Dehradun for Friday, under ₹800."

Two different machines could read that line. Watch what each one does with it.

Same request, two machinesone gives you words, one goes round a loop until the job is done

Try: press ▶ run the agent. Watch the dot go round: look, think, act. When it asks you something, you decide.

A chatbot
Get me a train ticket from Delhi to Dehradun for Friday, under ₹800.
Sure! Trains between Delhi and Dehradun include the Doon Shatabdi, the Shivalik Jan Express and the Ganga Valley Mail. Fares start at about ₹285. You can check seats and book on the railway website.

And there it stops. It gave you words. You still have to search, compare, check seats and pay.

An agent
laps 0 · tools used 0 · asked you 0

    Look at what just happened. Inside the agent there is the same kind of model as inside the chatbot. It reads words and writes words. The agent is different only because the model sits inside a loop:

    • ①Look. Read what the world says right now: the list of trains, the seat status, your reply.
    • ②Think. Decide the one next move. Not the whole plan, just the next move.
    • ③Act. Do it, using a tool: search trains, check seats, ask you, pay. Then go back to look, because the world has changed.

    The waitlist is the moment that matters. The cheapest train looked perfect, and the agent only found out it was waitlist 34 because it acted and then looked again. A chatbot never looks again. It has no second chance to be right.

    And look where it stopped to ask you: just before paying. Searching can be undone. Paying cannot. Knowing when to stop and ask is part of being a good agent, and Section 4 turns it into a simple calculation.

    The realization

    agent = model + tools + a loop

    The model chooses. The tools touch the world. The loop lets it see what its last move did, and choose again. Everything else in this unit, and in the units after it, is about making that loop go right.

    In one sentence: an agent is a model that keeps going round look → think → act, so it can see the result of each move and change its mind, until the job is done or it needs you.

    Free preview · Unit 22 of 22

    That was section 1. The rest of the unit opens when you unlock it.

    8 more sections and the practice arena — 9 widgets, 7 checks and 8 solved problems in the whole unit (this preview had 1 widget and 0 checks).

    Unlock Unit 22

    1. 2

      How it picks its next move

    2. 3

      The chain that breaks

    3. 4

      Ask or act?

    4. 5

      Memory: the desk and the shelf

    5. 6

      Planning: paths multiply

    6. 7

      Can it trust its own checker?

    7. 8

      The planted instruction

    8. 9

      The Agent Playground

    9. 10

      Practice arena — eight problems, solved in full

    Unlock this unit for ₹299, or all nine paid units for ₹999 — one-time payment, full refund within 7 days. See pricing.