same work, less waste
Domain 1 · 17% of the exam
The city that
builds itself.
Nobody hires an architect to buy bricks. They hire one because a thousand bricks must become a place where life works. This episode is about turning a business problem into a living system — and knowing which kind of building to raise.
Objectives
What you’ll be able to do
- Translate a business problem into a Claude-based solution
- Design end-to-end architectures with feedback loops
- Select workflow, agentic, or augmented-LLM patterns
- Design multi-agent systems and orchestration strategies
- Decompose complex problems into buildable parts
- Align every decision to a business value pillar
The story
Meera and the mayor
The mayor of a drowning city calls Meera, an architect. "Claims are flooding in after the storm," he says — he runs an insurance company, and the city is a metaphor, but stay with us. "Forty thousand claims. My people read each one, find the policy, check the damage photos, and decide. It takes eleven days. I need it in one."
A junior engineer would hear make the reading faster and buy a bigger printer. Meera hears something else: an input (claims, photos, policies), a processing heart (understand, retrieve, judge), an output (a decision with reasons), and — the part everyone forgets — a feedback loop: every human override teaches the system where it was wrong.
Before she draws anything, she asks the mayor one question: "When this works, what changes for you?" — "Adjusters stop doing paperwork and start handling the hard cases." That is the value pillar the whole city gets built on. Not tokens. Not models. Reclaimed judgment.
Metaphor map
Images that stick
| Concept | Image | Why it sticks |
|---|---|---|
| Business problem | The commission | Nobody wants a building; they want what happens inside it. Start from the outcome. |
| End-to-end architecture | The water system | In, through, out — and sensors that report back upstream. Cities die without the return pipe. |
| Architectural patterns | Building types | Factory, free-roaming crane, library annex. Choosing the building is the architect’s first real decision. |
| Multi-agent system | The construction crew | A foreman decomposes the job; specialists execute; nobody carries the whole blueprint alone. |
| Decomposition | City blocks | You can’t build a city in one pour. Blocks are testable, replaceable, and independently owned. |
| Value pillars | Why the city exists | Efficiency, transformation, productivity, cost, SLAs — the tax base that justifies every crane. |
Pattern explorer
Three shapes, one decision
Workflow
You define the steps; the model executes inside them. Extract, then validate, then decide — the route never changes. Deterministic, testable stage by stage, and fully auditable.
- Use when
- the task is well-understood, repeatable, and the steps are known in advance
- Trade-off
- rigid — novel cases fall off the line and need an escalation path
- Exam cue
- words like “fixed steps”, “consistent process”, “audit every stage” → workflow
Agentic
The model plans its own route: it chooses which tool to call, reads the result, and decides what to do next, looping until the goal is met. Powerful for open-ended problems — and harder to bound.
- Use when
- the path can’t be scripted: research, investigation, multi-step problem solving
- Trade-off
- less predictable cost, latency, and behavior — needs guardrails and budgets
- Exam cue
- “dynamic”, “unknown steps”, “model decides which tool” → agentic
Augmented LLM
A single enriched model call: retrieval, a tool result, or structured context bolted onto one request. No loops, no orchestration. The simplest pattern that works — and the right default.
- Use when
- one question, one answer — Q&A, summarization, classification with context
- Trade-off
- no multi-step reasoning across tools; outgrows itself if tasks compound
- Exam cue
- “single call”, “retrieval + generate”, “simplest solution” → augmented LLM
Value pillars
Multi-agent orchestration
When one worker can't hold the whole problem, appoint a foreman. An orchestrator decomposes the job, routes subtasks to specialists, and assembles the result — the construction crew, not the lone genius.
new capabilities, new business
experts on expert work
measurable spend reduction
speed & reliability, contractual
In production
Meera's claims triage, deployed
In production · insurance
Meera's claims triage, deployed
- PATTERN
- Workflow for the 80% routine path; agentic escalation lane for complex claims needing investigation.
- DECOMPOSITION
- Intake → damage classification → policy retrieval → coverage judgment → payout draft. Each step separately testable.
- FEEDBACK
- Adjuster overrides logged as labeled examples; weekly review feeds the evaluation set and prompt revisions.
- VALUE
- Routine claims: 11 days → same-day. Adjusters redeployed to complex cases — the productivity pillar, measured.
Storyboard
The film, shot by shot
01
8s
Every AI system begins as a commission — a problem wearing a deadline.
Cameraaerial dolly-in over a dark city at dawn, one office window lit
Motionrain particles; a phone glows; title type tracks in letter by letter
Soundlow rain bed, distant thunder, phone buzz
02
10s
The amateur hears “faster.” The architect hears input, processing, output… and the part everyone forgets.
Cameratop-down on a blueprint table, slow 25° orbit
Motionsticky notes morph into three glowing system blocks; a fourth arc draws itself in orange
Soundpaper slides, soft synth pulse per block
03
12s
Think of it as water. Claims pour in. The plant understands, retrieves, judges. Decisions flow out.
Cameraside-scroll following a glowing droplet through pipes
Motionliquid light travels input→plant→taps; pipe sections illuminate as it passes
Soundwater flow, rising harmonic as the plant lights up
04
8s
And when a human says “no, wrong” — that correction flows back upstream. The city learns.
Camerareverse tracking shot, camera pulls back along the return pipe
Motionred override spark turns amber, travels backwards, plant recalibrates with a shimmer
Soundreverse whoosh, single warm chime on arrival
05
12s
Three buildings. The assembly line. The free-roaming crane. The library annex. Choose the simplest that works.
Camerasplit-screen triptych, each panel a slow push-in
Motionfactory pistons in rhythm; crane pivots freely between sites; a shelf rolls into a reading room
Soundthree distinct motifs: metronome, wind, page-turn
06
10s
And when one worker isn’t enough — appoint a foreman. Decompose. Orchestrate. Build the city that builds itself.
Cameracrane shot rising from foreman to full city skyline, stars resolve into a node graph
Motionwork packets fan out to specialist crews and return assembled; skyline lights cascade
Soundorchestral swell, cut to silence on title card
Flashcards
Recall drill
Select a card to reveal the answer.
Quiz
Exam drill
Question 1
A retail chain processes supplier invoices: extract line items, validate against purchase orders, post to the ERP. The steps never vary and finance requires an audit trail for each stage. Which pattern fits best?
Fixed, repeatable steps with per-stage auditability are the signature of the workflow pattern. Agentic flexibility (A) adds unpredictability with no benefit; one call (C) can’t give stage-level audit; debate (D) addresses none of the requirements.
Question 2
An internal analyst tool must answer open-ended competitive-intelligence questions: search filings, query a database, cross-check the results, and sometimes write code to chart them. The steps differ per question. What should you recommend?
The path is unknown in advance and tool choice depends on intermediate results — the defining case for agentic architecture. A workflow (A) can’t enumerate the paths; stuffing context (B) fails on freshness and scale; fine-tuning (D) doesn’t grant tool access.
Question 3
An executive sponsor asks why the claims-triage system justifies its budget. Adjusters now spend their time on complex cases instead of routine paperwork, and routine claims settle same-day. Which value pillar does this map to most directly?
Reclaiming expert time for higher-value work is the productivity pillar. No new business line exists (B); same-day settlement is an outcome, not a contractual SLA (C); no reduction in headcount was stated (D).
Revision
Close your eyes and remember
A commission, not a purchase. Water through the city: in, through, out — and back. Pick the simplest building that works. Earn every agent. Trace every decision to a pillar.
Domain 1 · 17% of your exam · the largest single domain after Integration