Introduction to Agentic AI and AI Agents · Section 2 · Notes

Notes: Understand How AI Agents Work (Models, Tools and Memory)

Read a lecture instead of watching it, then make each idea yours.

Arun Nagarathanam · 29 Oct 2026 Notes · 5 of 10 in this course

Lecture 2.0

What Is Inside an AI Agent? (Why the Shirt Still Needs Folding)

When an AI answers your question but leaves the action undone, something is missing, and this section helps you say what. By the end you can look at a run in Claude Code or Codex and name which part did what.

Key ideas

01The webinar job: The agent read the planning-meeting transcript, prepared corrected invitation slides for Omar to review and saved a short update for the manager.
02Which part did what: You will say which part chose the next step, which carried out a search and which wrote the answer.
03What the app keeps: You will also see how the application keeps your requirements, and what its searches found, available while the work continues.

After this section you can look at a run and say:

which part chose the next step
which part carried out a search
which part wrote the answer

Knowing which part does what tells you what is missing when AI answers your question but leaves the action undone.

Lecture 2.1

Why Can AI Explain a Job Without Doing It? (Models and Actions)

An AI can describe how to fold a shirt while the shirt stays on the table. The part that works out the answer and the part that can act are not the same thing.

Key ideas

01Model: The part that interprets your request and works out a response. Sonnet and Opus are models inside Claude.
02App: Claude, ChatGPT and Gemini are apps. The app gives you a way to work with a model and decides which files and actions the model can use.
03Action needs access: To edit a spreadsheet, the app must open the file and you must allow the change.
04Switching models: A different model does not add that access. The new model may also hand you a formula to copy.

The shirt. Weave Robotics' Isaac 0 is a machine that actually folds it. A chat can only tell you how.

For a spreadsheet, getting the formula into the file takes three things:

the formula entered in the right cell
the file saved
an app that can open the file, plus your permission to change it

Switching to a different model does not fix a missing action. If the app cannot open and change the file, you are still copying the formula yourself.

Ask for the action. 'I want you to put the formula in the spreadsheet and save the updated file.' asks for the result, not only the answer.

Lecture 2.2

How Does an AI Agent Take Action? (Tools and Connectors)

An appointment written in a reply is not an appointment in your calendar. This lecture shows what a connection to Google Calendar adds, and why you still check the calendar afterwards.

Key ideas

01Connector: What Claude calls a connection to a service. Depending on the connection and permissions you choose, it can let the AI use your email, appointments or team messages.
02Tool: Each action is separate: get_event reads an event, create_event creates one and update_event changes an existing one.
03Plugin catalog: Codex offers connections this way. In the recording, Codex said it could not create the appointment from that conversation, even though the plugin lists the action.
04Account permission: The connected account must also allow changes. A company can allow reading appointments while blocking changes.

The supplier call. Claude writes the call for Friday at 3 p.m. in its reply, but the call is not in Google Calendar. To create it, Claude needs a connection that lets it create an event.

Before asking an agent to book something, check:

that the conversation can use the tool
that the connected account allows changes

Afterwards, open the calendar and check that the event is there at the right date and time.

Lecture 2.2A

How to Get the Most From This Course (Your Learning Path)

Each section comes with resources, and a few habits make them work harder for you. This is how I suggest using them.

Key ideas

01In order: Follow the course in order the first time, because later explanations use ideas the earlier lectures introduce.
02Summaries: Each section has a downloadable summary with the main ideas and key words in one place.
03Prepared files: Practical lessons come with prepared files, so you spend your time on how the models handle the job.
04Quizzes: Answer before you read the explanation. A surprising answer marks an idea worth another look.

Here is how I would work through each section:

watch the lectures in order, using the titles to find an explanation again
open the prepared files for the practical lessons
where an assignment has a sample answer, use it to judge your own result
take the quiz and answer before reading the explanation

In this section, the model comparison project gives you sample invoices and business files.

Lecture 2.3

What Is an API? (How Software Requests Information and Actions)

An API is how one program asks another for information or an action. Follow the supplier call from ordinary words to an event that Google Calendar creates and sends back.

Key ideas

01API: Short for application programming interface. Google publishes instructions for these requests: which actions exist, what each one needs and what comes back.
02Information or action: Asking for Friday's appointments returns events already there. Asking to create the supplier call adds a new event.
03Access: You grant it, for example by signing in and approving it. After that the connection sends the authorized requests for you.

The calendar tool needs these details:

the date
a 3 p.m. start and 3:30 p.m. end
a title
which calendar to use, if you have more than one

Another program sends those details to Google Calendar as a request. If the details and permissions are valid, Google Calendar creates the event and sends back the saved event's details.

That reply is what lets you check the result.

Lecture 2.4

How Does an AI Agent Work? (Look, Think, Do, Check)

Every agent run can be read as four moves. Watch Claude Code turn a project-management question into searches and a table, then check one row against the provider's own page.

Key ideas

01Look: The information the agent used, such as the returned search results.
02Think: The choice it made, such as which applications and plans to investigate.
03Do: The action: sending 5 WebSearch queries and producing the table.
04Check: Comparing the result with its source, such as opening the provider's pricing page.

The request asked for three things:

3 project-management applications under 20 US dollars per user per month
which ones let a client join as a guest without a paid account
a source link under every number

Claude Code sent 2 broad searches first, then specific searches for ClickUp Unlimited, Asana Starter and Trello Standard.

Your check. In the recording, Asana Starter showed US$10.99, matching the table. Billing and guest conditions still needed separate checks.

Links make the check possible. They do not show the agent performed it.

When a result leads the agent to choose another action and keep going, that repetition is often called an agent loop.

Make it yours

Freelancer or consultant

Your input
A choice you are comparing, such as tools or suppliers, with the price limit.

Ask AI
Compare these options under my limit and put a source link under every number.

Check
Open one source link and match the plan name and price.
Small business owner

Your input
The tool or supplier question you are deciding, with your price limit.

Ask AI
Which options fit my limit and my condition? Show a source link for each number.

Check
One price on the provider's own page.
On a team at work

Your input
The requirements your team needs, such as guest access without a paid account.

Ask AI
Find the plans that meet them, with a source link for each.

Check
Billing and guest conditions on the provider's page, since the table may not cover them.

Lecture 2.5

Can You Spot an Agent's 4 Moves in Your Work? (Agent Workflows)

The four moves are not only for AI. Daniel's unpaid-invoice reminders show them in ordinary work, and the same moves tell you where an AI could help.

Key ideas

01Look: Daniel reads the payment records and each client's messages, checking for any payment or question about the bill.
02Think: The firm's rules decide the response for each client.
03Do: Writing the 2 messages.
04Check: Comparing the amounts with the payment records and the wording with the reminder rules, before sending.

The firm's rules lead to two different messages:

a polite reminder to the client who usually pays late and is 9 days overdue
a firmer request for a payment update to the new client who is 30 days late with no payment or dispute

An AI could prepare the drafts if it had the invoice records, client messages and company rules.

Daniel still checks. He reads them before anyone sends them, and checks for payments again later.

The handout has more examples, such as a customer return and a supplier order. Each names the original record you use to check the result.

Make it yours

Freelancer or consultant

Your input
Your invoice list and the messages clients have sent about each.

Ask AI
Draft a reminder for each overdue invoice, following my reminder rules.

Check
Each amount against the payment record.
Small business owner

Your input
Payment records, customer messages and your rules for late payment.

Ask AI
Sort who needs a polite reminder and who needs a firmer one, then draft both.

Check
Each message matches the rule for that client.
On a team at work

Your input
The records, messages and company rules your employer lets you use.

Ask AI
Prepare the drafts for review. Do not send anything.

Check
You read every draft before anyone sends it.

Lecture 2.6

What Does an AI Agent Remember? (Context Window and Memory)

A model only knows what it is given for the reply it is writing now. See what Claude Code supplies for Daniel's client reply, what can drop out when space runs low and where a lasting rule belongs.

Key ideas

01Context: The information supplied to the model for the reply it is writing now, such as the client's message and the company's instruction.
02Context window: How much the model can work with at once, measured in tokens (a whole word or part of one). It holds what is supplied and the answer it writes.
03Compaction: Claude Code can shorten earlier conversation into a summary when space runs low. A summary can drop a detail such as an approval rule.
04CLAUDE.md: The project's instruction file that Claude Code reads when it starts work. A rule for future replies goes here.

The reply. A client asks for the next report as a spreadsheet instead of a PDF. The firm's rule says to keep replies courteous and not promise a change until the operations team approves it.

Typing only 'Write a reply to this client' tells the model neither. Claude Code reads the client's message and the company instruction and supplies them with Daniel's request.

When the window fills, three things follow:

Claude Code can compact earlier conversation into a summary
a summary can drop a detail, such as the approval rule
the original files stay where they are, so Claude Code can read them again

A lasting rule. 'Keep client replies courteous and do not promise a change before approval.' goes in CLAUDE.md. Keeping information for later use is commonly called memory.

Lecture 2.7

What Is an LLM in an AI Agent?

A fluent answer is not proof that it is true. This lecture explains what a large language model does and why you check a fact before you act on it.

Key ideas

01LLM: Sonnet and Opus learned patterns from large amounts of text during training. Models that work with language at this scale are called large language models.
02One type of model: The word model also covers systems built for other jobs, such as predicting demand for a product. Some language models also work with images or audio.
03A token at a time: After 'Thank you for your', it might write 'message' or 'payment'.
04Hallucination: A model can invent a price or feature and present it as true. This kind of error is often called a hallucination.

How it writes. An LLM writes its response a token at a time, using what it learned and the current context.

Why you check. A convincing sentence does not prove a claim. For a software price, compare the answer with the provider's pricing page.

Lecture 2.8

What Can AI Read and Create? (Text, Images, Audio and Video)

Before choosing a model, ask what information the job needs and what result you want back. That one question decides whether a model can do the work at all.

Key ideas

01Input and output: The information you give the AI is the input. The result you ask for is the output.
02Transcription: Turning speech into text, done by a speech-to-text model. Hugging Face lists it as 'automatic speech recognition'.
03OpenRouter: Lists models from different providers and the input types they support, with filters for text, image, audio and video.
04Listed is not running: A listing does not mean the model runs in your app.

Match the input to the job:

A receipt total. The input must include the picture. The app can pass the image to a model that reads images, or convert its words to text first, and either way the amount must survive.
A PDF table. Its layout must be kept so each amount stays with its item.
A meeting summary. If it depends only on the words said, turn the speech into text first.
Someone's tone of voice. The model needs the audio itself.

Creating pictures, speech, music or video is a separate output question.

Lecture 2.9

How Do You Choose an AI Model? (Benchmarks, Cost and Speed)

Choosing a model comes down to three things: how well it handles your work, what using it costs and how long you wait. Benchmarks help you shortlist, and your own job decides.

Key ideas

01Benchmark: A defined test used to compare systems. Artificial Analysis runs tests on models, and its Intelligence Index combines several into one score.
02Price per 1 million tokens: API prices are often quoted separately for input tokens (your request and material) and output tokens (what the model writes).
03Subscription: A monthly price with its own usage limits. Check your plan's usage information.
04The wait: Artificial Analysis reports time to the first answer token and output speed in tokens per second.

Menus change. Model menus change with updates. In the recording, Codex showed model names without descriptions, and Claude Code described only one model on hover.

Read the score carefully. Look at which benchmarks are included before reading the Intelligence Index.

A lower price per token does not always mean a cheaper job. These add up too:

long answers
repeated requests
extra reasoning tokens
the time you spend correcting the answer

Measure the whole wait. Reasoning tokens can arrive before the answer starts, so measure the full response time, and count waits for websites and tools too.

Lecture 2.10

How Can You Compare AI Models? (Artificial Analysis, OpenRouter and Arena)

The useful model is the one that handles your job well. Here is a fair comparison of two models on the same invoice review, with the correct answer worked out first.

Key ideas

01Same job, same checks: Give both models the same folder, request and checks, in 2 fresh Claude Code conversations with comparable settings.
02Candidates, not winners: Artificial Analysis and OpenRouter help you choose candidates.
03Arena: Shows 2 answers with the model names hidden and turns votes into rankings. A preferred answer can still hold a factual error.
04Record the run: Note each run's time and what you would correct, and try a few more examples before relying on a choice.

The correct answer first. For the four invoices:

Harbor & Pine owes 1,800 dollars of 2,400 and gets a courteous reminder that acknowledges the 600 paid
Northline is 30 days overdue with no payment or dispute, so it gets a firmer request for a payment update
Willow Ledger has paid in full and needs no reminder
Seaglass is disputed, so it goes to a person and gets no reminder

The result. In the recording, Sonnet and Opus both met every check. The usage breakdown showed Sonnet's cost a little lower.

If both models pass, compare waiting, corrections and the usage the app reports for each conversation.

Make it yours

Freelancer or consultant

Your input
A folder of real files from one of your jobs, and the answer you already know is right.

Ask AI
The same request in two fresh conversations, one on each model.

Check
Each model against your known answer, then time and corrections.
Small business owner

Your input
The invoices or orders you review each month, with your own rules.

Ask AI
Look at all the files in this folder, work out what to do with each, and prepare what is needed.

Check
The paid and disputed ones are handled the way you would handle them.
On a team at work

Your input
A prepared sample folder your employer lets you use.

Ask AI
Run the same request on two models and keep each run's results.

Check
The usage the app reports for each conversation.

Lecture 2.11

What Changes When You Switch AI Models? (Instructions and Memory)

Switching models inside one app usually keeps your setup. Moving to a different app is a separate check, so know what to test before you do.

Key ideas

01Stays in the folder: The invoice files and reminder policy stay in the project folder. Claude Code reads them and gives them to whichever model you choose.
02The app manages it: Inside the same app, saved preferences and connections such as Google Calendar may still be available. The model uses what the app supplies.
03Still test: A new model can produce a different result from the same files and instructions.
04Different app: Moving from Claude Code to Codex may need the calendar connected again and the reminder-policy file supplied.

When you buy an AI product for your company, ask the seller what happens when its model is replaced:

do your saved rules still apply?
do connections still work?
who tests the tasks you rely on?

Lecture 2.12

What Can You Now Explain About an Agent Run? (Section 2 Recap)

You can now take an agent run apart and say what each piece did. This recap puts the section's parts together, and Section 3 starts with ChatGPT.

Key ideas

01Model and search tools: In the project-management comparison, the model chose which plans to investigate and search tools carried out the requests.
02API: Gives software a way to request an action or information. A listed tool still has to be available in the task, with the access the action needs.
03Context and saved rules: Claude Code supplies current information and saved instructions as context within the model's window. An omitted rule can be supplied again from its file.
04Checks that stay: The invoice checks stay the same when the model changes.

The project-management run. The returned information helped the model build the table. Opening a provider's page compared a claim with its source.

The invoice checks that do not change with the model:

Harbor & Pine has 1,800 dollars outstanding
the paid invoice needs no reminder
the disputed invoice needs a person

Looking ahead. To judge an AI product, you also need to know how the product uses these parts to choose its actions. Section 3 starts with ChatGPT.

Quick reference

The parts of an agent

Model	Interprets your request and works out a response. Sonnet and Opus are models inside Claude.
App	Claude, ChatGPT, Gemini, Claude Code and Codex. Decides which files and actions the model can use.
Tool	A separate action, such as get_event, create_event or update_event.
Connector	What Claude calls a connection to a service such as Google Calendar. Codex offers them through its plugin catalog.
API	The agreed way programs request information or actions from a service.

Look, Think, Do, Check

Look	The information the agent used, such as returned search results.
Think	The choice it made, such as which plans to investigate.
Do	The action, such as sending searches and producing a table.
Check	Comparing the result with its source, such as the provider's pricing page.
Agent loop	A result leads the agent to choose another action, and the cycle repeats.

Memory and context

Context	The information supplied for the reply being written now.
Context window	How much the model can work with at once, measured in tokens.
Token	A whole word or part of one.
Compaction	Earlier conversation shortened into a summary. It can drop a detail.
CLAUDE.md	The project instruction file Claude Code reads when it starts. A lasting rule goes here.
Memory	Keeping information for later use.

Models and choosing

LLM	A large language model. It writes a token at a time using learned patterns and the current context.
Hallucination	A made-up fact presented as true. Compare facts you act on with the source page.
Input and output	What you give the AI and the result you ask for.
Benchmark	A defined test used to compare systems. Artificial Analysis combines several into its Intelligence Index.
API price	Quoted per 1 million tokens, often separately for input and output.
Arena	Two answers with model names hidden. A preferred answer can still contain an error.

Requests from the lectures

Ask for the action	I want you to put the formula in the spreadsheet and save the updated file.
Saved instruction	Keep client replies courteous and do not promise a change before approval.
First model	Look at all the files in this folder. Work out what we should do with each invoice, and prepare any reminders that are needed.
Second model	Look at all the files in this folder (exclude 'model comparison results' folder). Work out what we should do with each invoice, and prepare any reminders that are needed.