Solution Design and Architecture: CCAR-P domain 1 study guide

CCAR-P · Solution Design & Architecture (17% of the exam)

Domain 1 is Solution Design & Architecture, 17% of the CCAR-P exam. It tests one decision above all: given a business problem, choose the simplest Claude architecture that meets the requirement, and defend that choice against a more complex option.

What the official guide covers

The Claude Certified Architect Professional exam guide (version 1.0, effective July 2026) lists six tasks under Domain 1, "Solution Design & Architecture":

What the guide lists	What it means in practice
Translate business problems into Claude-based solutions	Turn "speed up claims handling" into inputs, outputs and success measures
Design end-to-end architectures, from input to processing to output to feedback loops	Draw the whole path, including how corrections flow back
Select architectural patterns (workflow, agentic, augmented LLM)	Decide who controls the next step: your code or Claude
Design multi-agent systems and orchestration strategies	Decide when subagents beat one agent, and what each receives
Apply decomposition techniques	Split a complex job into steps you can test alone
Align solutions to business value pillars (efficiency, transformation, productivity, cost, performance SLAs)	Tie each choice to an outcome the sponsor can measure

From business problem to design

Answer five questions before you pick a model or a pattern:

What output does the business need, and who uses it?
What are the inputs, and how often do they change?
How will you know an output is correct? This becomes the test set.
What does a wrong output cost, and who catches it?
What volume and response time must it handle?

Questions 3 and 4 shape the design more than model choice. An output that moves money or changes a record needs a check in code or a person.

Split the work: Claude, existing systems or people

Before you choose a pattern, give every step of the request one owner. Four model behaviours drive the split: outputs vary between runs, the context window is a hard limit, a wrong answer sounds as sure as a right one, and knowledge of rare, private or changing facts is unreliable.

Owner	Takes the steps that	Example: an IT access-request assistant
Claude	Need language: reading, classifying free text, drafting	Work out which system and access level the employee is asking for
Existing systems	Already have a reliable answer: rules engines, databases, live records	Check the role's entitlements in the identity system; grant standard access
People	Need judgement or accountability, or cannot be undone	Approve access to payroll or production data

Some steps split: Claude drafts the reply to the employee, the ticketing system sends it. Test each assignment with three questions: can a wrong call be undone, what does it cost, and who must answer for it?

The common mistake is moving a fixed business rule into the prompt because one component looks simpler. "Orders over 10,000 need credit approval" works on clean numbers, then fails silently when the amount arrives as "roughly ten thousand" in a sentence, and no log shows the misroute. Let Claude extract the amount; let code apply the threshold.

The end-to-end shape

Name all four stages in every design document.

Stage	What it holds	Questions to answer
Input	Requests, documents, events, system data	Is it trusted? Does it need validation or redaction?
Processing	Prompts, retrieval, tool calls	Single call, workflow or agent? Which model tier per step?
Output	Data, text, actions in other systems	Is it validated before anything acts on it?
Feedback loop	Corrections, reviewer decisions, logs	How do failures become test cases and prompt changes?

The feedback loop is the stage candidates forget. When two options differ only in whether they feed corrections back, the one that does is usually right.

Workflow, agent or augmented LLM?

Anthropic's "Building effective agents" defines the terms. An augmented LLM is a model call with retrieval, tools and memory. A workflow runs LLM calls and tools "through predefined code paths". An agent lets Claude direct its own process and tool use. Start simple; add agentic steps only when simpler solutions fall short.

Situation	Choose	Why
One question answered from known documents	Augmented LLM: one call with retrieval	Cheapest, fastest and easiest to test
Fixed steps in a known order	Workflow: prompt chaining	Each step tested and gated in code
Distinct request types	Workflow: routing	Each type gets its own prompt or model
Independent subtasks, or several views of one input	Workflow: parallelisation	Faster, or more reliable by comparing
Subtasks cannot be predicted in advance	Orchestrator-workers	A coordinator decides the subtasks at run time
Clear quality criteria, value from revising	Evaluator-optimizer	One call drafts, another critiques
Open-ended task with an unknown number of steps	Agent	Claude chooses the next step from what it has learned

When two patterns both look possible, check five factors in order and stop at the first that rules one out: can the steps be listed in advance, what a wrong answer costs, whether operations can reconstruct what happened, the latency budget, and cost per request at real volume. An agent is weakest on all five, so it needs a real reason. If an agent's traces show only a handful of recurring paths, a router with one chain per path does the same job and can be audited.

The CCAR-F pages cover the mechanics: 1.1 Agentic loops, 1.4 Workflow enforcement and handoff and 1.6 Task decomposition.

When to use several agents

Anthropic's write-up of its multi-agent research system names good fits for a coordinator with subagents: work that splits into independent parts, material larger than one context window, and parts that need different tools. Tasks where agents must share context or depend on each other, including most coding tasks, fit poorly. Multi-agent runs also use far more tokens than a chat, so the task's value must cover the cost.

Each subagent gets a full brief. In the Claude Agent SDK a subagent sees only the prompt it is given, not the parent's conversation. The brief needs the objective, output format, sources or tools to use and the limits of the task.
Each subagent gets only the tools it needs. A reviewer gets read-only tools.
Check coverage before synthesis. The coordinator must confirm one result per unit it sent out. A subagent that timed out or returned nothing is retried or reported as a gap, never left out of a summary that reads as complete.
Protect the coordinator. A failed subagent can be retried; a coordinator that loses its state usually loses the run. Checkpoint its progress and pass one trace ID to every subagent.
Gate irreversible actions. A subagent that would archive, pay or delete waits for human approval; lower-stakes actions are sampled instead of each being approved.
Cap the tree. Subagents can spawn their own subagents. The Agent SDK lets you limit nesting depth, concurrency and spend (max_budget_usd).

For coordinator patterns and context passing in depth, see 1.2 Coordinator and subagent patterns and 1.3 Subagent invocation and context passing.

Tie the design to business value

Agree a measure for each relevant pillar with the sponsor before you build.

Pillar	What to measure	Design lever
Efficiency	Handling time per case	Automate repeated steps; people handle exceptions
Productivity	Time to first draft	Put Claude where people wait or retype
Transformation	A service that was not possible before	Redesign the process, not a faster copy
Cost	Cost per completed task, including review time	Model tier per step, caching, batch jobs
Performance SLAs	Response time, accuracy on the agreed test set	Pattern choice, parallel steps

Two short statements go with the design into the statement of work. The feasibility verdict is one of three: feasible as scoped, feasible with constraints (each one written down, such as a maximum document length or a review gate), or not feasible (name the constraint that rules it out and the scope cut that would change it). The value statement compares the measured baseline with the projected state in the same unit, subtracts run cost, and gives a payback period plus what happens if volume or gain per task comes in lower. If a person still reviews outputs, keep their time in the projection.

Example: an architecture decision record

ADR-007  Claims triage architecture
Status:   Accepted

Context
  4,000 claims a day arrive by email. Handlers spend most of their time sorting
  them into five types and checking coverage. Target: triage within 10 minutes.
  A "not covered" decision must never reach a customer without human review.

Options
  1. One agent with all tools that decides its own steps
  2. Workflow: route by claim type > extract fields > check coverage >
     human review for every "not covered" or "unclear" result
  3. Coordinator agent with one subagent per claim type

Decision
  Option 2.

Reasons
  Steps are known and fixed. Each can be tested alone. The human review must
  happen every time, which code enforces and an agent's choices do not.
  Lowest cost per claim.

Consequences
  A new claim type needs a new route. Claims that fit no route go to a person.
  Revisit if more than 10% of claims fall outside the routes.

Value pillars
  Efficiency: handler minutes per claim. Performance SLA: 10-minute triage.

Feedback loop
  Handler overrides are logged and added to the test set every week.

Rules that decide exam answers

The simplest pattern that meets the requirement wins. Known steps call for a workflow; one call with retrieval beats both when it is enough.
A rule or step that must always hold belongs in code. Thresholds, eligibility rules, approvals and validations go in the workflow, an existing system or a hook, not in a prompt or an agent's instructions. Live records such as a customer's tier come from the system of record.
Subagents only know what you tell them. When subagents duplicate work or miss the point, fix the brief before changing the model.
Several agents need parallel, independent work. Tightly linked steps do better in one agent or a workflow.
No design is complete without a feedback loop. Corrections and failures must reach the test set and the prompts.
Agree the value measure before the pattern. Start from a measurable outcome, not a technology.

Where it appears in the exam

Solution Design & Architecture carries 17% of the CCAR-P exam, second only to Integration. Expect a business situation (a volume, a response-time target, a regulated process) and a question about which pattern, decomposition or orchestration fits. The guide names financial services, healthcare, retail, technology, education and government as typical candidate industries.

Two sample questions

These are original Timo practice questions. They are not official exam questions.

Question 1

A bank's finance team produces a monthly reconciliation report. Each month the same steps run: pull totals from three systems, compare them, and write commentary for every variance above a set threshold. A developer proposes an autonomous agent with access to all three systems. What should the architect recommend?

A. A coordinator agent with one subagent per source system, so each system is explored in parallel.B. A single agent with all tools and a system prompt that lists the steps in order.C. An evaluator-optimizer loop that drafts and critiques the whole report until it passes.D. A workflow in code that fetches and compares the totals, then calls Claude once per variance to write the commentary.

Answer: D. The steps are fixed, so a workflow is cheaper, easier to test and predictable, and code does the arithmetic exactly. A adds orchestration cost for work that needs no exploration, B leaves a fixed process to the model's choices, and C refines the whole report when only the commentary needs Claude.

Question 2

A coordinator agent sends subagents to review competitors' public filings for a client. The summaries ignore the client's question and repeat each other. Each delegation says only "Research company X". What should the architect change first?

A. Give every subagent a tool that reads the coordinator's full conversation history.B. Rewrite each delegation to state the client's question, the output format, the sources to use and what the other subagents are covering.C. Move every subagent to the most capable model tier.D. Replace the subagents with one agent that researches each company in turn.

Answer: B. Subagents know only their brief, so a complete brief fixes both the missed question and the overlap. A floods each subagent with context it does not need, C does not supply the missing information, and D gives up the parallel work that suits this task.

Build exercise

Pick a slow process at your workplace and answer the five design questions above.
Draw the four stages (input, processing, output, feedback loop) for it and mark where a wrong output would cause harm.
Write a one-page decision record comparing a single augmented call, a workflow and an agent. Choose one, with reasons and consequences.
Name one value pillar and the measure you would report after the first month.

Practise this topic

Claude Certified Architect Professional practice exam: free, 20 questions, no sign-up
Claude Certified Architect hub
CCAR-P study guide: all topics
Worked example: Production Claude tool loop
Next topic: Models, Prompting and Context

Sources

Claude Certified Architect, Professional Exam Guide, version 1.0, effective July 2026 (Anthropic), Domain 1: Solution Design & Architecture
Anthropic: Building effective agents
Anthropic: How we built our multi-agent research system
Claude Agent SDK documentation: Subagents in the SDK