Agentic AI — Hype vs Real Engineering Value
Notes on Lecture 01 of Andrew Ng’s Agentic AI course: separating a popular label from what a system can reliably do.
Source
These notes come from the first lecture of Andrew Ng’s Agentic AI course, Introduction to Agentic AI. I have kept two things separate on purpose: what the lecture says, and what I make of it. Everything after the next section is my own working, not a summary of the lecture.
Agentic AI — Lecture 01: Introduction to Agentic AI
What the lecture says
Andrew Ng observes that the term agentic AI has become widely used in marketing, and is sometimes attached to almost everything involving AI.
He also points out that, beyond the hype, genuinely valuable applications are being built using agentic workflows.
Both statements are true at the same time. The word is overused, and there is real work underneath it. The useful skill is telling the two apart.
Two words worth keeping apart
These are the working definitions I use in this notebook:
Hype — exaggerated promises, marketing and excitement around a technology.
Engineering value — the measurable ability of a technology to solve real problems reliably.
Hype is about what is said. Engineering value is about what can be checked. A demo, a launch post or a popular label sits on the left. A task, a test set and a failure rate sit on the right.
Hype
Promises and claims
- What it is said to do
- How it is marketed
- How exciting it feels
Engineering value
Reliability, measurable results and real-world impact
- Reliability — does it work every time it should?
- Measurable results — can the outcome be counted?
- Real-world impact — did a real problem get solved?
What “agentic” refers to — my sketch
To judge a claim about agentic AI, it helps to have a plain picture of what the word is meant to describe. My current understanding: instead of asking a model for one answer in one step, the work is broken into several steps, some of which use tools, with a check before the result is accepted.
- Plan Break the task into smaller steps.
- Act Do a step, using a tool where needed: search, a database, code.
- Check Compare the result with what was asked.
- Revise Fix what failed the check.
Repeat until the result passes the check, or hand it to a person.
Seen this way, “agentic” describes how a system is put together, not how impressive it is. Whether a particular agentic system is valuable is a separate question, and it is answered by measurement.
Real-world examples
The examples below are my own illustrations, not taken from the lecture. Each one shows the same idea described twice: once as hype, once as an engineering claim that can be tested. The difference is rarely the technology. It is whether the task is specific and the result is checked.
1. Customer support email
A common place where agentic workflows are tried: replying to routine customer messages.
Hype version
“An AI agent that replaces your support team.”
Engineering version
For order-status emails, look up the order, draft a reply with the correct delivery date, and let a person approve it before it is sent.
How I would check it
- How many drafts are sent without edits?
- How often is the wrong order looked up?
- What happens when the order is not found?
2. Processing invoices
Documents arrive in many formats, and the same few fields have to be read from each one.
Hype version
“Fully autonomous finance operations.”
Engineering version
Read the supplier, amount and due date from each invoice PDF, record them, and flag any field the system is unsure about for a person to confirm.
How I would check it
- Field-by-field accuracy against a hand-checked sample
- Share of invoices flagged for review
- Errors that slipped through unflagged
3. Fixing a failing test
Software has an advantage here: the test suite is an automatic check.
Hype version
“An agent that builds whole applications by itself.”
Engineering version
Given one failing test, propose a change, run the test suite, and open the change for human review only if every test passes.
How I would check it
- Does the full suite pass, not just the one test?
- How many proposed changes do reviewers accept?
- Does the change stay small and on topic?
4. Writing a research summary
Searching, reading and drafting are separate steps, which makes this a natural multi-step task.
Hype version
“An AI that researches anything for you.”
Engineering version
Search for sources on one question, read them, and write a short summary in which every statement links to the source it came from.
How I would check it
- Can every sentence be traced to a cited source?
- Do the sources say what the summary claims?
- Were important sources missed?
How I test a claim
When I meet the word “agentic” in a product page, a paper or a talk, these are the steps I now go through before believing it.
- Claim Write down what is being promised.
- Task Name the exact task, input and expected output.
- Test Try it on real cases, including awkward ones.
- Measure Count successes, failures and how failures are caught.
- Decide Keep what is reliable; drop what only demos well.
My takeaway
Key takeawayThe popularity of a technology does not prove its usefulness. Its engineering value comes from reliably solving real problems.
This is my interpretation, not a quotation from the lecture. It is the question I want to keep asking through the rest of this course: not “is this agentic?”, but “does this reliably solve the problem, and how do we know?”