This one tests whether you can run an engagement, not just design a system. A sixth of it is evaluation, and a seventh is talking to stakeholders. Take it after Foundations, once you have delivered something from discovery through to production.
Take it if
You lead Claude delivery for clients or internal teams and own the outcome.
Skip it if
You have not yet taken a system through discovery, deployment and operation. Nothing in the rules stops you, but the exam assumes it.
Before you start
Check yourself against this list first. If something is missing, close that gap before you pay for an exam.
Three or more years of systems architectureAnthropic recommends this. Junior engineers are not the target.
Six months of LLMs in productionA prototype does not count. You need to have run one.
An end-to-end delivery behind youDiscovery through to operations, including talking to stakeholders.
Your employer is in the Claude Partner NetworkOnly partner staff can register. Learn more
3+ years of systems architecture, 6+ months of LLMs in production
If you fail
Wait 14, then 30, then 90 days
What the exam covers
Solution Design & Architecture17%
Claude Models, Prompting & Context Engineering13%
Integration19%
Evaluation, Testing & Optimization16%
Governance, Safety & Risk Management14%
Stakeholder Communication & Lifecycle Management14%
Developer Productivity & Operational Enablement7%
What to focus on
Choosing a pattern: augmented LLM, workflow or agent, and defending that choice against cost and latency.
Integration is the largest domain: capability bloat, auth gaps, RAG chunking, retrieval matched to data shape, and MCP against API against agent-to-agent.
Progressive discovery against loading everything into context up front.
Evaluation you can defend: mixed-method test sets, A/B tests, and telling prompt failure from retrieval failure from model mismatch.
Compliance named in the blueprint: GDPR, HIPAA and FedRAMP, plus human-in-the-loop design.
Discovery, documenting decisions, and managing SLA expectations. Senior engineers often skip this and lose 14% of the exam.
Everything on one mapEvery topic and key idea in this certification.
mindmap
root((CCAR-P))
Solution design
Pattern choice
Business value pillars
Decomposition
Models, prompts and context
Model trade-offs
Reuse
Context budget
Integration
Capability bloat
Retrieval fit
Protocols
Progressive discovery
Evaluation
Metrics
Mixed test sets
Diagnosis
Governance, people and operations
Human in the loop
Compliance first
Expectation setting
Team enablement
Lesson 1 of 5, 17% of the exam
Solution design
The exam starts with a business problem, then asks you to choose how much machinery it deserves.
Pattern choice
Augmented LLM, workflow or agent. Choose the simplest that meets the requirement.
Business value pillars
Tie the design to efficiency, productivity, cost or performance SLAs, because that is how it gets funded.
Decomposition
Break complex problems into steps you can test on their own.
How much machinery does it need?Move down only when the simpler option fails.
flowchart TD
A["Business problem"] --> B{"One call enough?"}
B -->|Yes| C["Augmented LLM"]
B -->|No| D{"Are the steps known?"}
D -->|Yes| E["Workflow"]
D -->|No| F{"One domain?"}
F -->|Yes| G["Single agent"]
F -->|No| H["Multi-agent system"]
Common trap
Recommending multi-agent because the client asked for it. Your job is to fit the design to the problem.
Try it yourself
Take a real request and write one page: pattern chosen, two rejected, and why.
At this level the question is the trade-off, not the technique.
Model trade-offs
Route by complexity. Paying for the top tier on simple work wastes money, and the cheapest tier on hard work wastes trust.
Reuse
Caching, modular prompts and Skills stop you paying for the same instructions repeatedly.
Context budget
Treat tokens as a budget and decide what earns its place.
Routing by complexitySpend the budget where it changes the answer.
flowchart LR
A["Request"] --> B{"How hard is it?"}
B -->|Simple| C["Fast tier"]
B -->|Normal| D["Balanced tier"]
B -->|Hard| E["Strongest tier"]
C --> F{"Confident enough?"}
F -->|No| E
Common trap
Truncating needed content to save tokens when caching would have solved it.
Try it yourself
Design a routing rule for one workload and estimate its monthly cost at two volumes.
The largest domain, where security, retrieval and protocol choices meet.
Capability bloat
Audit what each agent can do and remove what the role does not need.
Retrieval fit
Match chunking and retrieval to the data shape. Tables, code and prose need different handling.
Protocols
MCP for tools and data, a direct API or CLI for simple calls, agent-to-agent for separate agents working together.
Progressive discovery
Load tools and context when needed, rather than everything up front.
Choosing a connectionPick by who else needs the integration.
flowchart TD
A{"What is being connected?"} -->|"Tools or data for many apps"| B["MCP"]
A -->|"One simple call"| C["API or CLI"]
A -->|"Independent agents working together"| D["Agent to agent"]
Common trap
Adding monitoring to a tool the agent should not have at all. Remove the capability first.
Try it yourself
Review one agent's tool list, remove what it does not need, and write down the auth flow for what remains.
Accuracy, latency, cost, safety and security, agreed before the build.
Mixed test sets
Real labeled cases, edge cases and adversarial inputs together.
Diagnosis
Separate prompt failure, retrieval failure, hallucination and model mismatch. Each has a different fix.
Why did quality drop?Find the failing layer before you change anything.
flowchart TD
A["Answers got worse"] --> B{"Is retrieved context right?"}
B -->|No| C["Fix ingestion, chunking or index"]
B -->|Yes| D{"Did the model change?"}
D -->|Yes| E["Run evals on both and pin the old one"]
D -->|No| F{"Did the prompt change?"}
F -->|Yes| G["Roll back and compare"]
F -->|No| H["Check data drift and new query types"]
Common trap
Blaming the model when retrieval changed. Check what the model actually received.
Try it yourself
Build a 30-case eval set, then change one thing and report the difference.
Over a third of the exam sits here, and engineers tend to under-prepare for it.
Human in the loop
Put review on irreversible, high-impact actions, not on everything.
Compliance first
GDPR, HIPAA and FedRAMP shape what is allowed, so settle them before design.
Expectation setting
Turn "100% accurate" into measurable targets with a review process.
Team enablement
Shared configuration improves everyone by default, without relying on training.
Where human review belongsReview by impact, not everywhere.
flowchart TD
A{"Can it be undone?"} -->|Yes| B{"High impact?"}
A -->|No| C["Human approves first"]
B -->|No| D["Automate and log"]
B -->|Yes| E["Automate, then sample for review"]
Common trap
Treating Stakeholder Communication as soft content. It carries the same weight as Governance.
Try it yourself
Write a one-page risk register and explain one trade-off as you would to an executive.
Exam facts checked on October 31, 2026 against CCAR-P exam guide, July 2026. Lessons are our own explanations, so use the official docs as the final word, and confirm price and dates before booking.
