PSG Polytechnic · 11 August 2026 · a 101
Three words you will keep hearing · what each one actually means, and when you need it
The Whole Hour, In One Picture
Each one is the previous one plus a single new idea. That is the entire progression.
Why · 8 min
Why
You are going to be asked to build an “AI agent” before anyone explains what one is. The word is doing a lot of work in job posts and almost none of it is precise.
So this hour is not about a framework or a vendor. It is about the three ideas underneath all of them, which will still be true after the tool names change.
You do not need to have written much code to follow this. If you have used a chatbot, you have used the first of the three.
Part 1 · 10 min
What It Is
Text goes in. Text comes out. That is the whole contract.
It has read a great deal and it is very good at continuing a sentence. It is not connected to anything. It cannot see your files, your database, or today’s date unless you paste them in.
What It Cannot Do
Both of them matter more than how clever the model is.
Ask it about your college timetable and it will tell you, correctly, that it has no access. Being right about what it cannot see is not the same as being useful.
One question, one answer, finished. It cannot look something up and then decide what to do about it.
Fix those two and you have not built a smarter model. You have built an agent.
Part 2 · 15 min
What An Agent Is
Follow the arrows. Each ingredient has a place in the picture.
A search, a database, a calculator. Now it can reach.
It goes round again with what it just learned.
It decides when it is done. You did not write that stopping point.
The Test
Who picks the next step?
Nobody. There is no next step.
You do, in code. You knew the order before it ran.
The model does, as it runs, based on what it finds.
Write this one down. It sorts every system you will ever be shown, whatever it is called.
The Part Nobody Says
If you can write the steps down, write them down. A for loop is cheaper, faster, easier to debug, and it will not invent a fourth step at two in the morning.
Use an agent when the next step depends on what you find. If you cannot name a step that changes based on the answer to the one before it, you want a workflow.
That flexibility is what you are buying. The rest of this hour is what it costs.
How
01
Get a free key from Google AI Studio. No card, no cloud project.
02
Send one question to the model and print the answer. That is the whole of step one, and it is worth doing on its own.
03
Give it exactly one tool — a function it can call. A search, or even just a list of your files.
04
Write the loop by hand: ask, let it call the tool, feed the result back, ask again. Do not reach for a framework yet.
Write the loop once yourself and every framework afterwards is just packaging you can already read.
Part 3 · 18 min
Why A Second One
One agent with tools already does a lot. So why add another?
One reads fifty documents and passes on a page. The next never carries the fifty.
The one that reads the database should not be the one that sends email.
Ten independent files, ten agents, one wall clock.
Notice what is missing: “so each one can specialise in a role.” That is the reason most people give, and it is the weakest one.
What Everyone Builds
A hard task goes to one agent. It does a mediocre job. So you hire it a team.
Four agents. On the run behind this talk it cost about five times the tokens and five times the wait — and the answer was no better than the single agent’s.
Why It Gets Worse
Every arrow is a place where the task gets retold in someone else’s words. The task survives. The understanding of it does not.
Brooks wrote this down in 1975 about people: adding more of them to a late project makes it later, because the lines between them grow faster than they do.
Agents are worse than people here, for one reason: they never say “wait, I don’t follow.”
On a stronger model the same chain cost eleven times a single agent, and when asked for three sentences it produced eighteen.
Nobody wrote anything wrong. The instruction was four handoffs back, and each agent was answering the one in front of it.
Where The Cost Is
Every boundary charges you something. A bigger model pays the same tolls.
How To Fix It
The same handoff, done two ways. Follow what happens to the original note.
Make each agent hand on a claim with the quote it came from, instead of a paragraph in its own words. Now you can check the answer without trusting anybody.
Five Shapes
Agents read and write a common evidence store.
PreferA policy picks the structure per task.
BestAgents argue to consensus in free text.
WorstA manager agent routes the work.
LossyA fixed sequence, every time.
BrittleThe bottom three are agents imitating a human meeting. That imitation is what you pay for.
The Counterintuitive One
Before you add an agent, ask one thing: can it be wrong in a way the others cannot?
A different model, a different tool, or a different source of evidence.
It can catch what the others structurally cannot. That is real signal.
The same model with a reworded prompt and a different job title.
Three agents built this way agree with each other because they are the same agent. You paid three times for one opinion.
Part 4 · 7 min
The Part That Is Not New
Everything so far was systems. This part was never about the tools.
You did it on purpose and you can say why.
You want to know how it works, not only that it worked.
You don’t cut the corner nobody would ever see.
You can hold a problem open before closing it.
A course can teach you methods. It cannot make you any of these, and neither can the machine: it takes instruction rather than intent, and it cannot sit in ambiguity — it resolves to the average, every time.
Take It With You
Build the smallest thing you can actually explain. Make it bigger only when something forces you to.
The working notebook. Every idea in this deck, as code you can run on a free AI Studio key.
colab.research.google.com/drive/1brSW2oJTzEi4_0bd2698pYjEQVsEfJSr?usp=sharing