Tell Agentic AI, Chatbots and Workflows Apart, Complete Resource Pack

Ideas to keep

1. 01ChatGPT can answer a question, search a website or carry out several actions. You need an explanation that fits the work you are looking at.
2. 02By the end of the section you can explain when ChatGPT is acting as an agent.
3. 03For any AI product, ask how its steps are chosen. Does it follow a sequence somebody prepared, or can the model choose another action after reading a result?
4. 04That question connects the name on the product to what the product does.

Try it (optional)

1. Think of one AI product someone at work has called an agent.
2. Write down what you have actually seen it do.
3. Keep that note. You will test it against the questions in the next lectures.

Files and links

The section starts with one question you can ask of any AI product: does it follow a sequence somebody prepared, or can the model choose another action after reading a result?

• Section 3 study guide (PDF)/downloads/agents/section-03/Section-3-Study-Guide.pdf

Ideas to keep

1. 01Asked for a current monthly price, ChatGPT could answer from what the model learned in training. Without checking the pricing page, it may give you an older price.
2. 02With web tools available, ChatGPT can search for the price instead and give you the source page, such as the Notion pricing page in the recording.
3. 03Notion's pricing page opens on yearly billing, so the page shows $20 beside ChatGPT's $24 monthly answer. Open the source and read which billing option it shows.
4. 04When the model chooses a next action from what it finds, and uses tools to continue toward your result, the work is agentic and the system doing it is an agent.
5. 05Finding the pages, reading the prices and preparing the answer form a workflow. When the model chooses and takes actions inside it, it is an agentic workflow.
6. 06Judge the task it performs, the tools it uses and the choices it makes, not the product name.

Try it (optional)

1. Ask ChatGPT, with web search on, for the current monthly price of a subscription you know.
2. Open the source page it gives you and compare the price and the billing period.
3. Note whether it searched, and whether it chose another page after the first one.

Files and links

The recording shows ChatGPT with web search on, answering a question about Notion's monthly Business price and citing the Notion pricing page. The page opens on yearly billing, which is why it shows a different figure from ChatGPT's monthly answer.

• ChatGPThttps://chatgpt.com/
• Notion pricing pagehttps://www.notion.com/pricing

Ideas to keep

1. 01A chat window tells you how you talk to the software. It does not tell you how the software chooses its steps.
2. 02A fixed workflow opens each saved page, reads the price, compares it with last week's price and writes the report. If a page has no price, it marks the price as missing. Someone prepared those steps before the job began.
3. 03An agent works on the same job with more freedom. In the recording, a saved page had lost its price, the agent reported Not confirmed, and the activity log showed it chose to search the supplier's website.
4. 04A long list of steps in a prompt does not tell you which kind of system carries them out. Both kinds can use web tools, and either can start from a chat.
5. 05They can work together: an agent finds the prices, then a fixed calculation works out how much each price changed.
6. 06More freedom helps when you cannot know every step in advance, and it also gives the agent more choices to get wrong. Prepared steps make mistakes too, so check both.

Try it (optional)

1. Pick a repeated job you do, such as checking prices or collecting figures.
2. Write the steps you could prepare in advance.
3. Mark the points where an agent might need to choose a different next action.
4. Before using any report, check that each price belongs to the right product and that each calculated change is correct.

Files and links

The agent in the recording is Claude Code in the Claude desktop app. It was given a list of tiles, last week's saved prices and links, then this request. One saved page was a practice page with its price removed, so the agent reported that price as not confirmed and searched the supplier's website. Before you use any price report, check that each price belongs to the right product and that each calculated change is correct.

• Claude Code overview (Anthropic documentation)https://code.claude.com/docs/en/overview
• Wickes Sigford Grey Matt Porcelain Wall and Floor Tile, 600 x 300 mmhttps://www.wickes.co.uk/Wickes-Sigford-Grey-Matt-Porcelain-Wall+Floor-Tile---600-x-300mm/p/308488

Ideas to keep

1. 01A manual start: you open an application and ask, so your request starts the job.
2. 02A scheduled start: a system that supports scheduled work runs the comparison every Monday at 9 a.m. Writing every Monday at 9 a.m. in a message does not set up a schedule.
3. 03An event start: a service configured to detect invoice emails starts the work as each one arrives.
4. 04The request, time or event that starts a job is its trigger. The trigger tells you when work begins, not how the steps are chosen.
5. 05Two Monday price checks share a scheduled trigger. The one where the model chooses another page to search is the agentic one.
6. 06A message saying A new invoice arrived is a notification. The work is done only when the invoice details are in your spreadsheet.

Try it (optional)

1. Choose one job you would like an agent to do.
2. Decide its trigger: your request, a scheduled time or an event.
3. Write the result you expect, and how you will check it happened.

Files and links

No outside links in this lecture. This job description gives you a start you can check and a result you can check separately.

Ideas to keep

1. 01A customer emails your online shop: Where is my order? An AI product that helps you prepare the reply is an assistant, because it helps you with your work.
2. 02The model looks up the order, finds the delivery is late, and chooses to check the delivery company's latest tracking update before writing. The system doing that is an agent, and its way of working is agentic.
3. 03Agentic AI means AI choosing and taking actions toward a goal.
4. 04If the software prepares a reply whenever a customer email arrives, that is automation. An automated process can follow fixed steps or use an agent.
5. 05Reading an order, preparing a message and sending it are separate actions. You can allow the product to prepare replies and still require your approval before it sends one.

Try it (optional)

1. Take one job of your own.
2. Name the result you want.
3. Name one choice the model might make along the way.
4. Name the tool action needed to carry out that choice.

Files and links

To try the pattern on your own job, name the result, one choice the model might make, and the tool action needed to carry it out.

• AI Agents course resources/resources/library/agents/
• quiz-3.4.md/downloads/agents/section-03/quiz-3.4.md

Ideas to keep

1. 01In a meeting you rarely need a definition. You need a sentence that helps people decide.
2. 02The word agentic is useful when you give the action and the reason: it chose another search because the first page did not answer the question.
3. 03Agent-washing is the gap between what an advertisement promises and what the demonstration shows.
4. 04Questions about connectors, tools and APIs ask what a connection can actually do, such as create a calendar event or accept a cancellation request.
5. 05Separate how the work proceeds from what starts it: prepared steps or model choices, and which trigger.
6. 06All twelve sentences are on the next sheet, with the lecture where each idea was explained.

Try it (optional)

1. Read the twelve sentences on the next sheet.
2. Pick two that fit a discussion you expect at work.
3. Put them into your own words, keeping the meaning.

Files and links

The twelve terms, each with a sentence you can use and the lecture where it was explained, are on their own page.

• Meeting vocabulary companion/resources/library/agents/section-03-meeting-vocabulary/

Ideas to keep

1. 01A prepared workflow could look up the order and draft a reply with fixed steps.
2. 02In the agentic version, the model found the delivery was late and chose an extra tracking check before writing.
3. 03A tool supplied the current information. Sending stayed a separate action that needs permission.
4. 04The email's arrival could start the process automatically, or a person could start it. The trigger tells you when the job begins; the later choices tell you how it works.
5. 05Neither the word assistant nor a chat window settles those questions.

Try it (optional)

1. Go back to the product you noted in Lecture 3.0.
2. Describe its trigger, how its steps are chosen, and which actions need permission.
3. Check your description against the twelve meeting sentences.

Files and links

• Notes: Tell Agentic AI, Chatbots and Workflows Apart/resources/library/agents/section-03-summary/
• summary.md/downloads/agents/section-03/summary.md
• Section 3 study guide (PDF)/downloads/agents/section-03/Section-3-Study-Guide.pdf