An interactive story

The Case
for Agents

A detective story about AI orchestration

Plain LLM → Single Agent → Multi-Agent Orchestration

Press Next or use arrow keys →

The case

Elena's mother is sending money to a stranger.

Elena
"Three months. $60,000. To a man named Thomas Whitfield — an 'engineer working overseas' she met on a dating site. She's in love. We're terrified."
Elena
"She won't listen to me. I need to know if he's real — and I need proof she'll believe."

The question

Is Thomas Whitfield real — and is he running a scam?

Elena's first instinct: ask an AI.

Elena → AI Chat
"My mother is sending money to a man named Thomas Whitfield, an engineer working overseas. His photo looks too good. Is this a scam?"
🤖
AI (no tools)
"This does sound like a romance scam. Common indicators include: overseas worker stories, requests for money, reluctance to meet in person. I'd recommend being cautious and speaking with authorities."

No photo check. No database search. No document analysis. Just pattern-matched text — reflecting back what Elena already told it. Her mother won't believe this. Elena needs proof, not opinions.

A language model with no tools can only reason over what you give it. It can't verify a photo, search a scam database, or read EXIF metadata. Elena needs an agent — something that can act, not just respond.

Her next stop: a detective.

Plain LLM = one inference, no external actions, no verification. An agent adds a loop (perceive → reason → act → observe) and tools (search, vision, databases). The loop is what turns language into investigation.
Act 1 — The solo detective
Detective Miles
Solo investigator · 12 years experience
web search image search records
Miles
"Leave it with me."

He starts with the name. Watch what he actually does.

Miles — thinks
"Thomas Whitfield, engineer, overseas. Let me check public records first."
🔍
tool: web search
"Thomas Whitfield" + "engineer" + location keywords
Miles — observes
"No matching profiles. No LinkedIn. No social footprint. Dead end."
Miles — adapts
"Name's fabricated or too common. Try the photo Elena brought."
🖼
tool: reverse image search
Uploads profile photo. Scans for original source.
⚠
Match found. Romanian model's Instagram, posted 2019. Photo is stolen.
significant finding
✓
Photo stolen — confirmed. Notes it. Moves on to documents.
closes thread 1, opens thread 2
Miles — thinks
"Photo's stolen. Probably same guy who ran that Lagos case last year. Anyway — onto the emails."
!
Missed connection. That photo appeared in 3 prior cases — same scammer network. Miles had the link but never cross-referenced it. Thread 1 is closed.
single agent weakness — no parallel memory, no cross-thread comparison

Why agents became a big deal

Before agents, LLMs could only reason over what you gave them. Miles changes that — he doesn't just answer, he acts. Searches the web. Runs image analysis. Queries databases. The loop that lets him do this — perceive, reason, act, observe, repeat — is what made AI agents genuinely useful for the first time.

Single agent runs threads sequentially. By the time thread 4 runs, thread 1 is cold — out of active reasoning. No second agent reads the photo finding fresh and compares it against the case database. Same data, missed connection.
Act 2 — Meanwhile

Day 3. Miles is on thread 2 of 5.

Miles · Day 3
✓ Photo stolen
✓ Email language
○ Photo metadata
○ Scam database
○ Case files
"Another week. Maybe ten days."
Same day — Nexus Agency
"I heard you solve cases in hours?"
"We work differently here. Case on my desk — I'll have answers today."

Miles isn't slow because he's bad at his job. He's sequential by design — one thread at a time, no parallelism, no specialization. For a case with 5 independent threads, that's 5× the time. The bottleneck is architectural.

One agent = one thread at a time. A multi-agent system with parallel specialists doesn't just work faster — it works differently. Same information, completely different result.
Act 3 — The agency
Chief
Orchestrator · Nexus Agency
decompose assign synthesize
Chief
"We work differently here. I read the case once. Then I tear it apart. Each thread goes to a specialist — someone with exactly the right tools. They work simultaneously. I pull it back together."
🔍
Aria
Web search · identity verification
📄
Jin
Document analysis · language patterns
🗄
Sam
Pattern research · scam databases
📷
Iris
Vision · photo forensics · metadata
🗂
Sage
Memory · past case files · precedent
This is multi-agent orchestration. Each agent is a specialist with its own tools and loop. The Chief (orchestrator) coordinates them — decompose, assign, collect, synthesize. Miles's 10-day case: delivered in 35 minutes.
Act 4 — Decomposition
Chief
"One owner per thread. Clear scope. Clear stopping condition. No overlap."
Chief (orchestrator) Aria web search identity A Jin documents analysis J Sam patterns databases S Iris photo forensics E Sage memory search Ar
Aria

Verify identity via public records and image search. Stop when confirmed or ruled out.

Jin

Read the wire transfer emails. Flag language patterns. Documents only — nothing else.

Sam

Cross-reference romance scam databases. Return count and closest structural matches.

Iris

Forensic photo analysis only. Metadata, compression artifacts, original upload source.

Sage

Pull similar cases from agency files. Same account structure, same transfer pattern.

Orchestrator = read → decompose → assign scope → define stopping conditions → dispatch. No agent touches another's thread. Duplication is impossible when ownership is explicit.
Act 5 — The investigation

They move at the same time.

Five threads. Five agents. Running simultaneously — the thing Miles could never do alone.

Aria
public records
image search

Photo stolen. Identity fake. ✓

Jin
reading docs
flagging patterns

Scam script. ✓

Sam
querying DB
matching

14 cases. ✓

Iris
photo forensics

Bucharest 2019. Re-uploaded. ✓

Sage
searching files

3 agency matches. ✓

~35 min
Nexus Agency · done
Day 3...
Miles · still on thread 3
Miles · at the same moment
"Alright, photo metadata checked. Now the scam database…"
Parallel execution: 5 agents × ~7 min each = 35 minutes total. Sequential = 5× longer. Same case, same information — completely different architecture.
Act 6 — Guardrails

Elena is desperate.

Elena
"Can you get into her phone? Pull her location? Access the bank account directly? I need everything."
Chief
"Stop. Not because of rules — because it would destroy your case."
→ Accessing private accounts without authorization is illegal. Evidence obtained this way is inadmissible.
→ Pulling her location tips off the scammer. He vanishes before anyone can act.
→ What we have from what you legally shared? That holds up in court.
Chief
"Instead: a documented evidence package — ready for the bank's fraud department and adult protective services. Today."
Tool permissions are explicit — agents can only call tools they've been granted access to. Guardrails aren't limitations on capability. They protect the mission, the user, and everyone affected by the answer.
Act 7 — The answer

Chief pulls the threads together.

Aria · Iris
Thomas Whitfield doesn't exist. Photo stolen from Romanian model, 2019.
Jin
Emails match textbook scam script. Urgency, secrecy, escalating requests.
Sam · Sage
14+ cases matched. 3 in agency files. One victim recovered funds.
Iris · Jin
EXIF metadata: photo re-uploaded from Bucharest, 2019. 3 artifacts confirm repurposing.
Internal Affairs · reads cold
"Sam's 14 cases — same actor confirmed, or same structural pattern? Not the same claim."
Chief
"Same pattern. Actor unconfirmed. Flagged in the report — we don't hide uncertainty."
Chief → Elena
Your mother is being scammed.
Evidence package ready. Take it to your bank's fraud department today. Contact adult protective services. Do not confront Thomas Whitfield — let institutions act. There is a path to recovering the funds.
Orchestrator decomposed, assigned scope, dispatched in parallel, synthesized.
Tool-calling agents reached outside the model — search, vision, documents, databases.
Memory agent surfaced what a blank-context agent would have missed.
Verification agent caught one overconfident claim before it reached Elena.
Guardrails kept the mission legal — and the evidence admissible.

Miles: 10 days, 2 threads. The Agency: 35 minutes, 5 threads, verified, admissible.

1 / 9
by Hui Wai Kit · May 2026