Understand How AI Agents Work (Models, Tools and Memory), Complete Resource Pack

Ideas to keep

1. 01In the 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.
2. 02By the end of this section you can look at a run in Claude Code or Codex and say which part chose the next step, which part carried out a search and which part wrote the answer.
3. 03You will also see how the application keeps your requirements, and the information its searches found, available while the work continues.
4. 04Knowing which part does what tells you what is missing when AI answers your question but leaves the action undone.

Try it (optional)

1. Think of one time an AI tool told you how to do something instead of doing it.
2. Keep that example in mind for the shirt-folding lecture that comes next.

Files and links

No outside links in this lecture.

• Section 2 study guide (PDF)/downloads/agents/section-02/Section-2-Study-Guide.pdf

Ideas to keep

1. 01An AI can describe how to fold a shirt while the shirt stays on the table. Weave Robotics' Isaac 0 is a machine that actually folds it.
2. 02When you send a message to Claude, the part that interprets your request and works out a response is called the model. Sonnet and Opus are models inside Claude.
3. 03Claude, ChatGPT and Gemini are apps. The app gives you a way to work with a model, and it decides which files and actions the model can use.
4. 04To edit a spreadsheet, the formula has to be entered in the right cell and the file saved. That needs the app to open the file, and it needs your permission to change it.
5. 05Switching to a different model does not fix this. If the app cannot open and change the file, the new model may also hand you a formula to copy.

Try it (optional)

1. Say exactly what you want done, including the change and the save.
2. Check whether the app has a way to open and change the file.
3. Check that you have allowed it to make the change.

Files and links

The laundry-folding machine in the lecture is Weave Robotics' Isaac 0. The request below is the one spoken in the lecture, for an app that can open and save your file.

• Isaac 0 (Weave Robotics)https://www.weaverobotics.com/isaac-0

Ideas to keep

1. 01Claude writes the supplier 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 to Google Calendar that lets it create an event.
2. 02Claude calls these connections connectors. They can let the AI use your email, appointments or team messages, depending on the connection and the permissions you choose.
3. 03Each action is a separate tool: get_event reads an event, create_event creates one and update_event changes an existing one.
4. 04Codex offers connections through its plugin catalog. In the recording, Codex said it could not create the appointment from that conversation, even though the plugin lists the action.
5. 05The connected account must also allow changes. A company can allow reading appointments while blocking changes.

Try it (optional)

1. Open the connection page and read the list of actions.
2. Sort them into actions that read information and actions that change something.
3. Before asking the agent to book the call, check that the conversation can use the event-creation tool.
4. Afterwards, open Google Calendar and check the call appears at the agreed date and time.

Files and links

The recording shows the Google Calendar connector in Claude's settings (its tools include get_event, create_event and update_event) and the Google Calendar plugin in Codex.

• Use connectors to extend Claude's capabilities (Claude Help Center)https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities
• Connectors and plugins (Claude Marketplace)https://claude.com/marketplace/connectors-plugins
• Plugins in Codex (OpenAI)https://learn.chatgpt.com/docs/plugins

Ideas to keep

1. 01Follow the course in order the first time, because later explanations use ideas the earlier lectures introduce. Lecture titles help you find one explanation again.
2. 02Each section has a downloadable summary with the main ideas and key words in one place.
3. 03Practical lessons come with prepared files. In this section, the model comparison project gives you sample invoices and business files, so you spend your time on how the models handle the job.
4. 04Where an assignment has a sample answer, use it to judge your own result.
5. 05The quizzes are a short way to check your understanding. Answer before you read the explanation; a surprising answer marks an idea worth another look.

Try it (optional)

1. Turn on captions, pause or change the playback speed whenever you need to.
2. Open this section's summary when you want a reminder of the key words.

Files and links

Every section has a summary, and practical lessons come with prepared files.

• AI Agents course resources/resources/library/agents/
• Section 2 notes/resources/library/agents/section-02-summary/

Ideas to keep

1. 01The calendar tool needs the details: the date, a 3 p.m. start and 3:30 p.m. end, a title, and which calendar to use if you have more than one.
2. 02Another program can send 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.
3. 03Google publishes instructions for these requests: which actions exist, what each one needs and what comes back. That agreed way for programs to ask for information or actions is the API, short for application programming interface.
4. 04Asking for Friday's appointments returns events already there. Asking to create the supplier call adds a new event.
5. 05You still grant access, for example by signing in and approving it. After that the connection sends the authorized requests for you.

Try it (optional)

1. Describe the appointment in ordinary words.
2. Check the returned event: is the date and time right, and is it in the right calendar?
3. If Google Calendar rejected the request, expect an error, not a claim that the call was booked.

Files and links

These are Google's published instructions for the Google Calendar API, including the request that creates an event.

• Google Calendar API overview (Google for Developers)https://developers.google.com/workspace/calendar/api/guides/overview
• Events: insert, the request that creates an event (Google for Developers)https://developers.google.com/workspace/calendar/api/v3/reference/events/insert

Ideas to keep

1. 01The request asked for 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, and a source link under every number.
2. 02Claude Code sent 2 broad searches first, then specific searches for ClickUp Unlimited, Asana Starter and Trello Standard. The returned results are Look.
3. 03Choosing which applications and plans to investigate is Think. Sending 5 WebSearch queries and producing the table is Do.
4. 04Opening a source link and comparing the plan name and price with the provider's page is your Check. In the recording, Asana Starter showed US$10.99, matching the table; billing and guest conditions still needed separate checks.
5. 05The links make the check possible. They do not show that 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.

Try it (optional)

1. Open one source link from the table.
2. Compare the plan name and price with the provider's page.
3. Check whether the price needs a full year's payment and whether a client can join as a free guest.
4. If a price does not match, ask the agent to look again; that starts another Look.

Files and links

These are the provider pages behind the comparison table in the recording. Prices on them change; the figures in the lecture are what the recording showed on the day it was made.

• Asana pricinghttps://asana.com/pricing
• Trello pricinghttps://trello.com/pricing
• ClickUp pricinghttps://clickup.com/pricing

Ideas to keep

1. 01Daniel reads the payment records and each client's messages, checking for any payment or question about the bill. That is Look.
2. 02The firm's rules call for a polite reminder to the client who usually pays late and is 9 days overdue, and a firmer request for a payment update to the new client who is 30 days late with no payment or dispute. Choosing the response is Think.
3. 03Writing the 2 messages is Do. Comparing the amounts with the payment records and the wording with the reminder rules, before sending, is Check.
4. 04An AI could prepare the drafts if it had the invoice records, client messages and company rules. Daniel still reads them before anyone sends them, and checks for payments again later.
5. 05The handout's examples, such as a customer return and a supplier order, each name the original record you use to check the result.

Try it (optional)

1. Pick one example from the handout.
2. Name the information used, the choice made, the action taken and the record that checks it.
3. Optionally, do the same for a recurring job of your own where the next action depends on what you find.

Files and links

The handout has Daniel's reminder and four more worked examples, each with the record you use to check the result.

• Handout: Look, Think, Do and Check examples/downloads/agents/section-02/2.5-look-think-do-check-examples.md
• Quiz: Trace the four moves in an agent's job/downloads/agents/section-02/quiz-2.5.md

Ideas to keep

1. 01A client asks for the next report as a spreadsheet instead of a PDF. The firm's rule says: keep replies courteous and do not promise a change until the operations team approves it.
2. 02Typing 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. That supplied information is the context.
3. 03The model can work with only so much at once; that capacity is its context window. Text is measured in tokens, a whole word or part of one. For these Claude models, the window holds what is supplied and the answer it writes.
4. 04When space runs low, Claude Code can shorten earlier conversation into a summary. This is compaction, and 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.
5. 05A rule for future replies can go in CLAUDE.md, the project's instruction file that Claude Code reads when it starts work. Keeping information for later use is commonly called memory.

Try it (optional)

1. In Claude Code, open the context usage display and see which categories use space.
2. If a reply ignores a rule, ask Claude Code to read the client's message and the current instruction again.
3. Open CLAUDE.md to read the saved rules, and update it when your company's rules change.

Files and links

• How Claude remembers your project (Claude Code Docs)https://code.claude.com/docs/en/memory
• Context windows (Claude Platform Docs)https://platform.claude.com/docs/en/build-with-claude/context-windows

Ideas to keep

1. 01Sonnet and Opus learned patterns from large amounts of text during training. Models that work with language at this scale are called large language models, or LLMs.
2. 02An LLM is one type of model. The word model also covers systems built for other jobs, such as predicting demand for a product. Some language models can also work with images or audio.
3. 03An LLM writes its response a token at a time, using what it learned and the current context. After 'Thank you for your', it might write 'message' or 'payment'.
4. 04A model can invent a price or feature and present it as true. That kind of error is often called a hallucination.
5. 05A convincing sentence does not prove a claim. For a software price, compare the answer with the provider's pricing page.

Try it (optional)

1. When an answer gives a fact you will act on, find the page it should come from.
2. Compare the claim with that page before you use it.

Files and links

No outside links in this lecture.

Ideas to keep

1. 01The information you give the AI is the input. The result you ask for is the output.
2. 02For the total on a photographed receipt, 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; either way the amount must survive. A PDF table also needs its layout kept so each amount stays with its item.
3. 03OpenRouter lists models from different providers and the input types they support, with filters for text, image, audio and video.
4. 04If a meeting summary depends only on the words said, turn the speech into text first; that is transcription, done by a speech-to-text model. To judge someone's tone of voice, the model needs the audio itself.
5. 05Creating pictures, speech, music or video is a separate output question. Hugging Face lists models by task and calls speech-to-text 'automatic speech recognition'. A listing does not mean the model runs in your app.

Try it (optional)

1. Write down the input the job needs and the output you want.
2. Find models that support that input and output.
3. Check that your application accepts the input and passes it to the model.

Files and links

The recording shows OpenRouter's model list filtered to image input, AuraScribe's speech-to-text model settings, and Hugging Face's task filters.

• OpenRouter models filtered to image inputhttps://openrouter.ai/models?input_modalities=image
• Hugging Face models: automatic speech recognitionhttps://huggingface.co/models?pipeline_tag=automatic-speech-recognition&sort=trending
• Hugging Face tasks: automatic speech recognition explainedhttps://huggingface.co/tasks/automatic-speech-recognition

Ideas to keep

1. 01Model menus change with updates. In the recording, Codex showed model names without descriptions, and Claude Code described only one model on hover.
2. 02A benchmark is a defined test used to compare systems. Artificial Analysis runs tests on models, and its Intelligence Index combines several into one score. Look at which benchmarks are included before reading the score.
3. 03API prices are quoted per 1 million tokens, often separately for input tokens (your request and material) and output tokens (what the model writes). A monthly subscription has its own price and usage limits; check your plan's usage information for those.
4. 04A lower price per token does not always mean a cheaper job. Long answers, repeated requests and extra reasoning tokens add up, and so does the time you spend correcting the answer.
5. 05Artificial Analysis reports time to the first answer token and output speed in tokens per second. Reasoning tokens can arrive before the answer starts, so measure the full response time, and count waits for websites and tools too.

Try it (optional)

1. Use published tests to choose 2 or 3 models worth trying.
2. Check whether a quoted price is for input tokens, output tokens or both.
3. Check your own plan's usage page for limits and reset times.
4. Judge the whole job: answer quality, cost or usage, and total time.

Files and links

Artificial Analysis publishes the Intelligence Index, API prices per million tokens, time to first answer token and output speed.

• Artificial Analysis: compare models (intelligence, price, speed)https://artificialanalysis.ai/models
• Artificial Analysis benchmarking methodologyhttps://artificialanalysis.ai/methodology
• Model configuration (Claude Code Docs)https://code.claude.com/docs/en/model-config

Ideas to keep

1. 01The useful model is the one that handles your job well. Give both models the same job, source material and checks, in 2 fresh Claude Code conversations with comparable settings.
2. 02Work out the correct result first. 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.
3. 03Willow Ledger has paid in full and needs no reminder. Seaglass is disputed, so it goes to a person and gets no reminder.
4. 04Artificial Analysis and OpenRouter help you choose candidates. Arena shows 2 answers with the model names hidden and turns votes into rankings; a preferred answer can still hold a factual error.
5. 05In the recording, Sonnet and Opus both met every check. The usage breakdown showed Sonnet's cost a little lower. Record each run's time and what you would correct, and try a few more examples before relying on a choice.

Try it (optional)

1. Download model-comparison-project.zip and unzip it.
2. Follow the practice sheet after this lecture for the full steps.
3. Save each complete response in the model comparison results folder.
4. Check both answers against the answer key that follows the practice sheet.

Files and links

The practice page has the steps, both requests with Copy buttons, and the answer key. Download the project folder below and unzip it before you start. Arena says conversations and votes may be shared with model providers or in public research data, so keep private client and company information out of it.

• Compare 2 models on the same invoice review/resources/library/agents/section-02-model-comparison/
• model-comparison-project.zip/downloads/agents/section-02/model-comparison-project.zip
• Compare 2 models on the same invoice review (markdown)/downloads/agents/section-02/model-comparison-instructions.md
• Invoice comparison answer key (markdown)/downloads/agents/section-02/model-comparison-answer-key.md
• Artificial Analysis: Claude Sonnet 5https://artificialanalysis.ai/models/claude-sonnet-5
• Artificial Analysis: Claude Opus 5https://artificialanalysis.ai/models/claude-opus-5
• OpenRouter: Claude Sonnet 5https://openrouter.ai/anthropic/claude-sonnet-5
• OpenRouter: Claude Opus 5https://openrouter.ai/anthropic/claude-opus-5
• Arenahttps://arena.ai/
• How to compare AI models in Battle Mode (Arena Help Center)https://help.arena.ai/articles/4489017547-how-to-use-battle-mode

Ideas to keep

1. 01The invoice files and reminder policy stay in the project folder. Claude Code reads them and gives them to whichever model you choose.
2. 02Inside the same app, saved preferences and account connections such as Google Calendar may still be available, because the app manages them. The model uses what the app supplies.
3. 03A new model can still produce a different result from the same files and instructions, so test it on the same job.
4. 04Moving from Claude Code to Codex needs a separate check. You may need to connect the calendar in Codex and give it the reminder-policy file.
5. 05If 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, and who tests the tasks you rely on?

Try it (optional)

1. After switching models, run the same job and check the balances and draft messages.
2. After moving apps, reconnect what you need and supply the instruction files again.
3. Test the action itself, such as creating an appointment, and check the saved event's date and time.

Files and links

No new outside links. The calendar connection links are under 2.2, and the instruction-file link is under 2.6.

• Quiz: Choose and change models without losing the job/downloads/agents/section-02/quiz-2.11.md

Ideas to keep

1. 01In the project-management comparison, the model chose which plans to investigate, search tools carried out the requests and the returned information helped the model build the table. Opening a provider's page compared a claim with its source.
2. 02An API 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.
3. 03Daniel's client reply showed why current information and saved instructions both matter. Claude Code supplies them as context within the model's context window, and an omitted rule can be supplied again from its file.
4. 04The invoice checks stay the same when the model changes: Harbor & Pine has 1,800 dollars outstanding, the paid invoice needs no reminder and the disputed invoice needs a person.
5. 05To judge an AI product, you also need to know how the product uses these parts to choose its actions. Section 3 starts with ChatGPT.

Try it (optional)

1. Open the Section 2 summary and read the key words once more.
2. Explain one agent run you watched in this section using model, tool, context and check.

Files and links

• Section 2 notes/resources/library/agents/section-02-summary/
• Section 2 summary (markdown)/downloads/agents/section-02/summary.md
• Section 2 study guide (PDF)/downloads/agents/section-02/Section-2-Study-Guide.pdf