Domain 1: Solution Design & Architecture — 17%

Topic 1.1: Translate Business Problems into Claude-Based AI Solutions

What You Need to Know for the Exam

This objective tests whether you can take a real business problem and translate it into an appropriate Claude-based solution.

The exam is unlikely to simply ask:

“What is an agent?”

Instead, you may receive a business scenario describing:

• a problem the organization wants to solve;
• existing processes;
• users and stakeholders;
• performance expectations;
• cost constraints;
• security or compliance requirements;
• a proposed architecture.

You will then need to determine:

• what the actual business requirement is;
• what role Claude should perform;
• whether Claude is even needed for every step;
• which architectural pattern best fits;
• which constraints drive that choice;
• whether the proposed solution is over-engineered or under-engineered.

The official CCAR-P blueprint includes this objective under Domain 1: Solution Design & Architecture, which represents 17% of the exam. Candidates are expected to translate business problems into Claude-based solutions and align designs with business value such as efficiency, transformation, productivity, cost, and performance SLAs.

---

1. The Core Exam Mindset

The most important principle is:

Start with the business problem, not with Claude.

Do not assume that every business requirement needs:

• an agent;
• multiple agents;
• RAG;
• the most powerful model;
• several Claude calls;
• complex orchestration.

The exam often rewards the simplest architecture that satisfies the actual requirement.

Think in this order:

Business Problem
       ↓
Business Outcome
       ↓
Task to be performed
       ↓
Constraints
       ↓
Claude's role
       ↓
Architecture pattern
       ↓
Controls and validation
       ↓
Measurable business result

---

2. A Simple Framework to Use in the Exam

Use the following five-step framework whenever you see a business scenario.

O-T-C-P-F

O — Outcome

What business result is required?

Examples:

• reduce processing time;
• reduce operating cost;
• increase employee productivity;
• improve customer response time;
• automate repetitive work;
• improve accuracy;
• improve scalability.

---

T — Task

What work actually needs to happen?

Examples:

• classify;
• extract;
• summarize;
• retrieve information;
• generate a response;
• analyze evidence;
• call tools;
• decide what to investigate next.

---

C — Constraints

What limits the solution?

Look for:

• latency;
• cost;
• accuracy;
• auditability;
• security;
• compliance;
• throughput;
• human-review requirements;
• data freshness.

---

P — Pattern

What is the simplest suitable architecture?

Possible answers include:

• single Claude call;
• augmented LLM;
• workflow;
• agent;
• parallel workflow;
• orchestrator-worker or multi-agent architecture.

---

F — Feedback

How will the organization know the solution is working?

Examples:

• response time;
• task completion rate;
• review time;
• cost per transaction;
• error rate;
• human approval rate;
• customer satisfaction;
• evaluation results.

---

3. Example: Converting a Business Request into an Architecture

A stakeholder says:

“We want an AI agent for customer support.”

This is not yet a sufficient requirement.

An architect should translate it further.

Business problem

Support representatives spend too much time searching multiple policy documents.

Business outcome

Reduce average ticket handling time by 30%.

Required task

1. Understand the customer’s question.
2. Find the relevant policy.
3. Draft a grounded response.

Constraints

• policies change weekly;
• support representative sends the final reply;
• response should be generated quickly;
• the system must not autonomously issue refunds.

Suitable architecture

Customer Question
       ↓
Retrieve relevant policy
       ↓
Claude generates grounded response
       ↓
Support representative reviews
       ↓
Final response

This is likely an augmented LLM / RAG solution.

A complex autonomous multi-agent system would provide little additional business value.

---

4. First Identify the Shape of the Task

Architecture selection becomes much easier when you identify the task type.

Type 1 — Simple Transformation

Example:

Classify incoming support emails into one of eight categories.

Email
  ↓
Claude
  ↓
Category

Likely pattern

Single Claude call

Use this when the task is:

• narrow;
• well-defined;
• usually completed in one reasoning step.

---

5. Knowledge-Grounded Tasks

Example:

Employees need answers based on HR policies that change frequently.

Employee Question
       ↓
Retrieve relevant policies
       ↓
Claude
       ↓
Grounded Answer

Likely pattern

Augmented LLM / RAG

Use this when Claude requires:

• organization-specific knowledge;
• current information;
• a large document corpus;
• authoritative source material.

Exam clue

Look for phrases such as:

• “knowledge changes frequently”;
• “must answer using company documents”;
• “authoritative source”;
• “current policies”;
• “grounded response.”

---

6. Predictable Multi-Step Tasks

Example:

Process a loan application by extracting fields, validating them, applying established policy rules, and preparing a review packet.

Application
     ↓
Extract Fields
     ↓
Validate
     ↓
Apply Policy Rules
     ↓
Generate Review Packet

Likely pattern

Workflow

Use a workflow when:

• the steps are known;
• the order is known;
• the process should remain consistent;
• success can be verified;
• business rules are well-defined.

The Practitioner Workbook highlights workflow vs. agent as one of the most important Domain 1 distinctions.

---

7. Dynamic and Open-Ended Tasks

Example:

Investigate why a production application failed.

Claude may need to:

Inspect Logs
      ↓
Discover DB Error
      ↓
Inspect Database Metrics
      ↓
Discover Connection Saturation
      ↓
Inspect Recent Deployment
      ↓
Form Root-Cause Hypothesis

The exact path cannot be predicted before the investigation starts.

Likely pattern

Agent

Use an agent when:

• the next step depends on what Claude discovers;
• the number of steps cannot be predicted;
• Claude must dynamically choose tools;
• the problem is exploratory.

---

8. The Most Important Exam Distinction: Workflow vs. Agent

This is one of the highest-value concepts for Domain 1.

Ask:

Can the sequence of steps reasonably be determined before execution begins?

If YES

Use a:

Workflow

If NO

And the next action genuinely depends on intermediate findings:

Agent

Remember:

Workflow
Developer controls the path.

Agent
Claude dynamically determines the path.

---

9. Quick Workflow vs. Agent Comparison

Requirement	Workflow	Agent
Steps known beforehand	✓
Fixed business process	✓
Highly repeatable	✓
Easy to validate	✓
Next step depends on discoveries		✓
Number of steps unknown		✓
Tool selection changes dynamically		✓
Open-ended investigation		✓

Exam Shortcut

Known path = Workflow Discovered path = Agent

---

10. Do Not Confuse Tool Use with Agentic Architecture

A common mistake is:

“Claude uses tools, therefore it must be an agent.”

Incorrect.

A workflow can also use tools.

Example:

Invoice
   ↓
Claude extracts values
   ↓
ERP API validates supplier
   ↓
Rule engine checks limits
   ↓
Claude generates explanation

The system uses tools, but the path is predetermined.

It is still a workflow.

---

11. Augmented LLM vs. Agent

Suppose the requirement says:

Retrieve the relevant product policy and produce one grounded answer.

The solution may simply be:

Question
   ↓
Retrieval
   ↓
Claude
   ↓
Answer

You probably do not need:

Planning Agent
      ↓
Retrieval Agent
      ↓
Reasoning Agent
      ↓
Writer Agent
      ↓
Review Agent

Exam lesson

Do not introduce agentic architecture when:

• the task is straightforward;
• there is no dynamic planning;
• one retrieval/generation cycle solves the requirement.

---

12. Use the Architecture Complexity Ladder

When reading an exam question, move from the simplest solution upward.

Can one Claude call solve it?
          ↓ No

Does Claude mainly need extra knowledge/tools?
          ↓ Yes

Augmented LLM
          ↓

Are several processing steps required?
          ↓ Yes

Can those steps be defined beforehand?

        Yes              No
         ↓                ↓
     Workflow           Agent
                           ↓
             Are dynamic independent
             subtasks/delegation needed?
                           ↓
                     Multi-Agent

Exam Rule

Increase architectural complexity only when the business problem requires it.

---

13. Business Constraints Can Decide the Architecture

A business requirement may contain one phrase that eliminates several architectures.

Learn to spot these clues.

Requirement wording	Architectural implication
“same sequence every time”	Workflow
“consistent and auditable”	Workflow / deterministic controls
“depends on what is discovered”	Agent
“cannot predict the required steps”	Agent
“current company knowledge”	RAG / augmented LLM
“independent analyses”	Parallelization
“different known request categories”	Routing
“strict latency target”	Reduce sequential model/tool calls
“cost is the primary objective”	Avoid unnecessary agents/model calls
“must always enforce”	Deterministic control outside the model
“repeatedly improve against clear criteria”	Evaluator-optimizer

---

14. Business Value Must Drive Architecture

The official blueprint expects solutions to align with business-value pillars such as:

• efficiency;
• transformation;
• productivity;
• cost;
• performance SLAs.

For exam questions, always identify the primary business driver.

---

Efficiency Example

Business problem:

Customer-service representatives spend too much time researching policies.

Architecture:

Ticket
  ↓
Retrieve Policy
  ↓
Claude Draft
  ↓
Human Review

Measure:

Average handling time

---

Productivity Example

Business problem:

Analysts spend three hours reviewing large reports.

Architecture:

Report
  ↓
Claude extracts and summarizes
  ↓
Analyst verifies

Measure:

Review time per report

---

Cost Example

Business problem:

Process millions of straightforward classifications economically.

Good solution:

• simple architecture;
• appropriate model;
• batching where suitable.

Poor solution:

• several agents;
• multiple sequential calls;
• highest-capability model for every stage.

Exam lesson

An architecture may be technically valid but still be wrong because it conflicts with the business objective.

---

15. Performance and SLA Requirements

Suppose the requirement says:

Response must complete within one second.

Be suspicious of:

Claude
   ↓
Agent
   ↓
Tool
   ↓
Claude
   ↓
Tool
   ↓
Claude
   ↓
Final Response

Every serial step adds latency.

Exam principle

A hard performance SLA can invalidate an otherwise technically capable architecture.

---

16. Decomposition

Complex business problems should usually be decomposed before selecting the implementation pattern.

Example:

Evaluate whether the company should acquire another business.

This may contain:

Acquisition Analysis
        |
  ┌─────┼─────────┬─────────┐
  ↓     ↓         ↓         ↓
Finance Market   Legal   Cyber Risk
  └─────┼─────────┴─────────┘
        ↓
     Synthesis

The architect should determine for each subtask:

• Does Claude perform it?
• Does it require retrieval?
• Is deterministic code needed?
• Does it require a tool?
• Can it run in parallel?
• Does a human need to validate it?

---

17. Decomposition Does Not Automatically Mean Multi-Agent

This is another important exam trap.

If the process is:

Extract
   ↓
Classify
   ↓
Summarize

you do not automatically need three agents.

It may simply be one workflow.

Use separate agents only when they provide a meaningful advantage such as:

• different tool access;
• context isolation;
• parallel work;
• specialized capabilities;
• independent verification;
• dynamic task delegation.

---

18. Fixed vs. Dynamic Decomposition

Fixed Decomposition

You already know the subtasks:

Document
   ↓
Extract
   ↓
Classify
   ↓
Summarize

Likely:

Workflow

---

Dynamic Decomposition

Claude discovers what work is required.

Example:

Analyze a software repository and implement a requested feature.

Claude may discover:

Database change required
API update required
UI update required
Tests required
Configuration update required

These subtasks were not fully known beforehand.

Potential architecture:

Orchestrator-worker / agentic decomposition

---

19. Sequential vs. Parallel Execution

Architecture should reflect dependencies.

Sequential

When B needs A’s result:

A → B → C → D

Run sequentially.

---

Parallel

When tasks are independent:

          ┌→ Financial ─────┐
          ├→ Legal ─────────┤
Input ────┼→ Market ────────┼→ Synthesis
          └→ Security ──────┘

Parallel execution may reduce total latency.

Exam clue

Look for:

• “independent”;
• “can be performed simultaneously”;
• “latency is important.”

---

20. Claude vs. Deterministic Code

Claude should not make every decision.

Good uses of Claude

Use Claude for:

• natural-language understanding;
• extraction from unstructured text;
• summarization;
• synthesis;
• reasoning over ambiguity;
• drafting;
• adaptive tool selection.

---

Good uses of deterministic code

Prefer code for:

• exact calculations;
• fixed eligibility rules;
• permission checks;
• mandatory restrictions;
• transaction integrity;
• deterministic validations.

---

Example

Instead of:

Claude
   ↓
Decides regulatory fee

prefer:

Claude extracts attributes
       ↓
Deterministic rule engine
       ↓
Fee calculation
       ↓
Claude explains result

Exam principle

Use Claude for reasoning and language; use deterministic controls for rules that must always hold.

---

21. Worked Example — Loan Underwriting

Scenario

A lender wants Claude to:

• extract information from applications;
• apply established policy rules;
• generate a decision packet.

Auditors require consistent processing.

Step 1 — Outcome

Increase processing efficiency while retaining consistent auditability.

Step 2 — Task

Extraction + policy processing + explanation.

Step 3 — Constraint

Consistency and auditability.

Step 4 — Pattern

Workflow.

Recommended architecture

Loan Application
       ↓
Claude Field Extraction
       ↓
Schema Validation
       ↓
Deterministic Policy Engine
       ↓
Claude Explanation
       ↓
Human Review

Why not an agent?

The process is already known.

Agent autonomy adds unnecessary variability.

---

22. Worked Example — Production Incident Investigation

Scenario

Claude must investigate incidents using:

• application logs;
• database metrics;
• deployment history;
• network information.

The next investigation step depends on previous findings.

Analysis

Outcome:

Reduce incident-resolution time.

Task:

Open-ended investigation.

Constraint:

Path cannot be predetermined.

Architecture

Engineer Request
       ↓
Claude Agent
       ↓
Select Diagnostic Tool
       ↓
Review Result
       ↓
Choose Next Action
       ↓
...
       ↓
Root-Cause Hypothesis
       ↓
Engineer Validation

Correct pattern

Agent

---

23. Worked Example — HR Policy Assistant

Scenario

Employees need answers from company policies that change regularly.

Analysis

Outcome:

Improve employee self-service.

Task:

Retrieve relevant policy and generate answer.

Constraint:

Knowledge must stay current.

Architecture

Employee Question
       ↓
Policy Retrieval
       ↓
Claude
       ↓
Grounded Answer + Source

Correct pattern

Augmented LLM / RAG

Why not agent?

There is no meaningful dynamic planning requirement.

---

24. Worked Example — Due-Diligence Report

Scenario

A report requires:

• financial analysis;
• legal analysis;
• market analysis;
• cybersecurity analysis.

Each analysis is independent.

Architecture

                 ┌→ Financial ──────┐
                 ├→ Legal ──────────┤
Request ─────────┼→ Market ─────────┼→ Final Synthesis
                 └→ Cybersecurity ──┘

Correct design principle

Parallelize independent work.

Do not execute the tasks sequentially unless dependencies require it.

---

25. Common Exam Traps

Trap 1 — “Agents are more advanced, therefore better.”

Wrong.

Architecture should be selected based on the business problem.

---

Trap 2 — “Tool use means agent.”

Wrong.

Workflows can also call tools.

---

Trap 3 — “Complex problem means multi-agent.”

Wrong.

First decompose the problem.

---

Trap 4 — “More flexibility is always better.”

Wrong.

Flexibility can increase:

• cost;
• latency;
• variability;
• testing complexity.

---

Trap 5 — “The most capable model should always be used.”

Wrong.

Model and architecture choices must consider:

• quality;
• cost;
• latency;
• workload complexity.

---

Trap 6 — Ignoring the stated business metric

If the requirement says:

Reduce processing cost by 40%.

An architecture that significantly increases model calls may be inappropriate even if it improves reasoning quality.

---

Trap 7 — Using Claude for deterministic rules

If a rule must always hold, use deterministic enforcement.

Do not rely entirely on prompting.

---

Trap 8 — Solving an architecture problem with prompt engineering

If the system violates a strict latency requirement because it has six sequential calls, rewriting the prompt does not solve the fundamental issue.

---

26. How to Approach These Questions During the Exam

When you see a long scenario, ask these five questions.

Question 1

What business outcome matters most?

---

Question 2

What work actually needs to happen?

---

Question 3

What is the strongest constraint?

Examples:

• latency;
• cost;
• consistency;
• freshness;
• auditability.

---

Question 4

Is the processing path known beforehand?

YES → Workflow

NO → Consider Agent

---

Question 5

What is the simplest architecture that satisfies all requirements?

This question can eliminate many attractive distractors.

---

27. High-Yield Revision Table

Scenario Signal	Think
One simple transformation	Single Claude call
Need current/private knowledge	Augmented LLM / RAG
Known sequence	Workflow
Unknown next step	Agent
Known independent subtasks	Parallelization
Dynamic unknown subtasks	Orchestrator-worker
Strong auditability	More deterministic control
Strict latency	Fewer serial calls
Cost-sensitive high volume	Simpler architecture / efficient processing
Hard business rule	Code / rule engine
Poor outputs are not being detected	Feedback/evaluation loop

---

28. What to Memorize

You should be able to answer these without notes:

1.

What is the most important difference between a workflow and an agent?

Answer: Whether the execution path is predefined or dynamically determined based on intermediate results.

2.

When should you use an augmented LLM?

Answer: When Claude needs additional knowledge or capabilities but does not require open-ended autonomous planning.

3.

Does decomposition automatically require multiple agents?

Answer: No.

4.

When should deterministic code be preferred?

Answer: For hard rules, permissions, calculations, transaction controls, and other requirements that must always hold.

5.

What should determine architecture complexity?

Answer: Business requirements and constraints, not technical sophistication.

---

29. Recommended Exercise 1 — Architecture Triage

For each scenario, write only four things:

Business Outcome:
Task:
Binding Constraint:
Architecture:

Scenario A

A legal department reviews 3,000 supplier contracts every month. Lawyers spend significant time locating unusual clauses. The organization wants to reduce first-pass review effort by 50%. Lawyers remain responsible for final decisions.

Suggested Answer

Business Outcome: Reduce legal review effort by 50%.

Task: Identify and summarize relevant clauses.

Binding Constraint: Final legal judgment remains human.

Architecture: Document-processing workflow using Claude for extraction/analysis with human review.

---

Scenario B

An operations engineer wants Claude to investigate production incidents. Depending on what logs reveal, Claude may inspect database metrics, deployments, network events, or other systems.

Suggested Answer

Business Outcome: Reduce incident-resolution time.

Task: Investigate evidence across multiple tools.

Binding Constraint: Investigation path cannot be predefined.

Architecture: Agent with controlled diagnostic-tool access.

---

30. Recommended Exercise 2 — Simplify the Architecture

Consider this design:

Employee Question
       ↓
Coordinator Agent
       ↓
Intent Agent
       ↓
Retrieval Agent
       ↓
Policy Agent
       ↓
Writer Agent
       ↓
Reviewer Agent
       ↓
Answer

Business requirement:

Employees need answers from an HR policy repository. Answers should arrive quickly and policies change weekly.

Your task

Identify which complexity is unnecessary.

Better architecture

Employee Question
       ↓
Retrieve Relevant Policy
       ↓
Claude
       ↓
Grounded Answer + Citation

Reason

The business problem requires fresh knowledge and grounded generation, not autonomous planning.

This exercise is particularly useful because CCAR-P questions frequently include plausible but unnecessarily complex designs.

---

31. Practice Questions

Question 1

A retailer wants Claude to process return requests. Every request follows the same procedure:

1. extract order information;
2. verify eligibility against established rules;
3. calculate the approved refund;
4. prepare a customer response.

The organization requires consistent and auditable processing.

Which architecture is MOST appropriate?

A. Autonomous agent that dynamically chooses every next step B. Multi-agent system with a specialist agent for each stage C. Workflow with predefined steps and deterministic policy enforcement D. Orchestrator-worker architecture

Correct Answer: C

Why?

The process is:

• known;
• predictable;
• sequential;
• easily validated.

That is the workflow pattern.

Why the others are wrong

A: Adds unnecessary autonomy.

B: Adds complexity without a demonstrated need.

D: Dynamic task decomposition is unnecessary because the subtasks are already known.

Exam Takeaway

Known and repeatable process → Workflow.

---

32. Practice Question 2

A cloud operations team wants Claude to investigate outages. Depending on initial telemetry, Claude may inspect logs, database metrics, deployment history, or network systems. The investigation path cannot be predicted beforehand.

Which architecture is MOST appropriate?

A. Fixed workflow containing every possible diagnostic step B. Single summarization prompt C. Agent that selects diagnostic tools based on intermediate results D. Retrieval-only augmented LLM

Correct Answer: C

Why?

The strongest clue is:

The investigation path cannot be predicted beforehand.

Claude must repeatedly:

Observe
  ↓
Reason
  ↓
Choose Action
  ↓
Observe New Result

That is agentic behavior.

Exam Takeaway

Next step depends on discoveries → Agent.

---

33. Practice Question 3

A company wants to reduce the cost of processing two million simple product descriptions each month. Each description requires the same one-step classification.

The proposed architecture uses five specialized agents and the most capable model for every step.

What is the PRIMARY problem?

A. Agents cannot perform classification B. Multi-agent architectures cannot support high volume C. The design is unnecessarily complex and conflicts with the cost objective D. Classification requires retrieval

Correct Answer: C

Why?

The business objective is cost reduction.

The task is simple and repeatable.

Multiple agents and unnecessary model calls increase:

• token consumption;
• latency;
• operational complexity.

Exam Takeaway

A technically capable architecture can still be wrong if it does not align with the business objective.

---

34. Practice Question 4

A healthcare organization builds this solution:

Clinical Notes
      ↓
Claude
      ↓
Structured Summary
      ↓
Doctor

After deployment, inaccurate summaries occasionally reach doctors. The system has no mechanism for measuring these failures or incorporating them into future testing.

What is the MOST important missing architectural component?

A. Another Claude agent B. Feedback and evaluation mechanism C. Larger context window D. MCP integration

Correct Answer: B

Why?

An end-to-end architecture should include:

Input
  ↓
Processing
  ↓
Output
  ↓
Feedback / Evaluation

Without feedback, quality problems may continue without systematic detection or improvement.

Exam Takeaway

Production architecture does not end at output.

---

35. Practice Question 5

A company uses Claude to prepare due-diligence reports. Every report requires independent:

• financial analysis;
• legal analysis;
• market analysis;
• cybersecurity analysis.

None of the analyses depends on another, and overall response time is important.

Which design is MOST appropriate?

A. Execute the four analyses sequentially B. Execute the independent analyses in parallel and aggregate their results C. Let an autonomous agent randomly determine the execution order D. Combine all sources into one very large prompt

Correct Answer: B

Why?

The tasks are explicitly:

• independent;
• decomposable;
• suitable for simultaneous execution.

Parallel execution reduces overall latency.

Exam Takeaway

Independent subtasks + latency concern → Parallelization.

---

36. Final Exam-Day Cheat Sheet

Remember this sequence:

BUSINESS PROBLEM
      ↓
What outcome matters?
      ↓
What task must Claude perform?
      ↓
What constraint controls the design?
      ↓
Is the execution path known?
      ↓
Choose simplest suitable pattern
      ↓
Add validation and feedback

And remember these mappings:

Single well-defined task
→ Single Claude call

Needs external/current knowledge
→ Augmented LLM / RAG

Known multi-step process
→ Workflow

Unknown/adaptive process
→ Agent

Known independent subtasks
→ Parallel workflow

Unknown dynamic subtasks
→ Orchestrator-worker / Multi-agent

Mandatory deterministic rule
→ Code / policy engine

Strict latency or cost constraint
→ Reduce unnecessary model calls

The One Rule to Remember

Do not choose the most sophisticated Claude architecture. Choose the simplest architecture that satisfies the business outcome and its constraints.

That is the architectural judgment this CCAR-P objective is primarily testing.