Your intuition is reasonable. As written, I don't think this example demonstrates why agents are necessary. In fact, I'd probably implement the described system as a sequential pipeline.
Here's why.
What he describes
The workflow is essentially:
Scrape properties
↓
Extract structured data
↓
Calculate ROI
↓
Filter ROI > 6%
↓
Gather neighborhood information
↓
Analyze risks
↓
Generate report
Nothing here requires autonomous planning or dynamic decision making.
You could implement this as:
for property in listings:
scrape()
calculate_roi()
if roi > 6:
enrich_with_neighborhood_data()
analyze()
generate_pdf()
Even if an LLM performs the "analyze()" step, it's still just one stage in a deterministic pipeline.
What actually makes something an "agent"?
People often use "agent" to mean "an LLM with a prompt." That's become common, but it's not a very useful distinction.
A stronger definition is:
An agent can decide what actions to take next, based on intermediate results, rather than simply executing a predefined sequence.
The important difference isn't multiple prompts.
It's control flow.
A pipeline:
A → B → C → D
An agent:
Think
↓
Should I search?
↓
Search
Think again
↓
Need more information?
↓
Call another tool
Think again
↓
Still uncertain?
↓
Try another strategy
The agent is choosing its own next steps.
Looking at each of his "agents"
Agent 1
Scrapes listings
Deterministic.
Pulls out price
Deterministic.
Calculates ROI
Pure calculation.
If ROI > 6%, forward.
Simple if statement.
Nothing here requires agency.
Agent 2
This part is slightly more interesting.
He says:
checks neighborhood stats
If that simply means:
- query crime API
- query school API
- query rent API
- summarize
that's still a pipeline.
Even if GPT writes the summary.
Where agents would become useful
Suppose instead Agent 2 behaved like this:
"Crime data is unavailable. I'll try city open data."
Then:
"Still missing. I'll estimate using adjacent ZIP codes."
Then:
"The rental estimate looks inconsistent with Redfin. I'll compare with Apartments.com."
Then:
"This area has unusually low taxes. I should investigate whether it's in a flood zone."
Now the execution isn't predetermined.
The LLM is deciding:
- which tool to call
- in what order
- when to stop
- what information is missing
- whether confidence is sufficient
That's agent behavior.
Another example
Imagine it notices:
ROI = 8%
Normally that's good.
But then it notices:
- school rating dropped sharply
- employer leaving city
- property taxes increasing
- vacancy rate rising
It decides:
"I need additional market information before recommending this."
So it performs additional searches that weren't explicitly programmed.
A pipeline can't spontaneously add new stages unless you explicitly code every possibility.
Or imagine negotiation between agents
Suppose:
Research agent:
"This property scores 82."
Risk agent:
"Flood risk is high."
Finance agent:
"Insurance costs erase the ROI."
Investment agent:
"Reject."
Now you have multiple specialized agents exchanging information until they converge on a decision.
That's much closer to a true multi-agent system.
His example could easily stay a pipeline
In fact, I might build it like this:
Property Scraper
↓
ROI Calculator
↓
Filter
↓
Neighborhood Data Collector
↓
LLM Risk Analyzer
↓
PDF Generator
Very robust.
Easy to test.
Cheap.
Predictable.
When I'd redesign it as agents
Suppose the goal changes from:
"Analyze Zillow listings."
to
"Find me the best investment opportunities in Atlanta under $500k."
Now the system may need to:
- search Zillow
- search Realtor
- search county tax records
- compare recent sales
- estimate renovations
- search rental demand
- verify zoning
- investigate HOA rules
- estimate insurance
- search news
- decide information is insufficient
- perform another search
- compare against previous candidates
- ask the user follow-up questions
No one knows beforehand exactly which tools will be needed for every property.
That's an agent problem.
A useful rule of thumb
Ask yourself:
Can I write the flowchart before seeing the data?
If yes, it's usually a workflow or pipeline.
A
↓
B
↓
C
↓
D
If no—because the system must decide what to do next based on what it discovers—then an agent architecture becomes valuable.
Observe
↓
Reason
↓
Choose tool
↓
Observe
↓
Reason again
↓
Choose different tool
↓
Repeat until goal achieved
My assessment of the example
I think the example oversells the need for agents. The described process is largely a sequential workflow with one or two LLM-powered analysis steps. Simply labeling those stages "Agent 1" and "Agent 2" doesn't make the architecture agentic.
A stronger justification would involve situations where the "Investment Advisor" must autonomously determine what additional information is needed, decide which tools or data sources to use, recover from missing or conflicting data, iterate until it reaches a confidence threshold, or coordinate with other specialized agents. Those are scenarios where agents offer capabilities that a fixed pipeline doesn't naturally provide.
