Claude Certified Architect — Foundations
Live · 120 questions
Claude Certified Architect — Foundations practice exams
2 sets, 120 questions, every option explained. The build track between Developer and Professional: agentic loops and multi-agent orchestration, tool and MCP design, Claude Code configuration, structured output, and context reliability. The exam draws four scenarios from a published bank of six, and its hard questions turn on two answers that are both defensible.
Form 1 is free, no account.
Get full access $39All 828 questions across the four certifications.
“I found this set of questions so useful, compared with others courses here in Udemy, this is by far the best. Thanks, I passed !”
- Items per form
- 60
- Fee
- $125 USD
- Pass mark
- 720 / 1000
- Valid for
- 12 months
Booking the real exam needs a partner company email. No partner? See how sponsorship works
Start here
Form 1 is free — all 60 questions
No account, no email. Every option explained as you go.
What the exam weights
D1 Agentic Architecture & Orchestration · 27% of the exam · 32 questions in the bank · 16 on the free form
The 2 sets
- Form 160 questions · full lengthFree · above
- Form 260 questions · blueprint-weighted
The question most people get wrong
run_query lets the billing subagent compose arbitrary queries against the billing database. Twice last quarter it returned invoice rows belonging to a different account whose contact name matched the caller's. Which two changes most directly address this?
- A
Replace run_query with a lookup_invoice tool that takes an invoice ID and rejects invoices outside the account in disputeCorrect
- B
Rewrite run_query's description so it warns that the tool is a last resort, for questions no specific billing tool covers
- C
Remove run_query from the billing subagent's tool set once lookup_invoice covers the lookups the role actually needsCorrect
- D
Add a PostToolUse hook that drops any returned row whose account id is not the account currently in dispute
- E
Use forced tool selection so lookup_order runs first on each billing turn and establishes the account in context
66.7% of 60 people got this wrong in the last 30 days (as of 2026-09-15). The most popular answer was AD, chosen by 5 of the 15 who answered, and every one of them was wrong.
The principle
Guardrails belong in the tool interface, not in the instructions wrapped around it. Narrowing a generic tool means both building the constrained replacement and retiring the general tool it replaces.
How to prepare for CCAR-F
- 1Sit Form 1, free and without an account. Too long for today? The quick round is ten questions, one per domain.
- 2Read the explanation under every option, the wrong ones most of all. The hard questions on this exam turn on why the tempting answer is wrong.
- 3Check the result page for your weakest domain and objective, then take the readiness check for a second reading against everyone else who sat it.
- 4Study the guide for that domain. The 9 guides below are grouped by what the exam weights.
A worked question, every option explained
The support coordinator runs with get_customer, lookup_order, process_refund and escalate_to_human in its allowedTools, and the billing-dispute subagent is defined and described. The coordinator never spawns it. Which entry must allowedTools also contain?
- A
"Invoke", added to the coordinator's allowedTools
Names the thing being spawned, which is how the configuration reads, and an allowedTools list does hold the names the coordinator may invoke. This is not one of them, so nothing changes.
- B
"Delegate", added to the coordinator's allowedTools
Reads as the action the coordinator is meant to take, and delegation is genuinely what is being configured. The list takes tool names rather than intents, and no tool carries this one.
- C
"Subagent", added to the coordinator's allowedTools
Matches the vocabulary the rest of the configuration uses for the child agent, so it is the name an architect reaches for from memory. The spawning tool is not named after its result.
- D
"Task", added to the coordinator's allowedToolsCorrect
Correct. Spawning runs through the Task tool, and a coordinator can invoke subagents only when "Task" appears in its allowedTools.
The principle
Subagent spawning is a tool call: the coordinator needs "Task" in allowedTools, whatever its prompt and the subagent's description say.
Written by Matthew Hartman, who holds the Claude Certified Architect — Professional credential (965 out of 1000). Independent practice material, not affiliated with or endorsed by Anthropic.
9 free guides
- Choosing an examWhich of the four should you take?CCAO-F, CCDV-F, CCAR-F and CCAR-P side by side - items, fee, domains, and who each one is for.8 min
- EligibilityCan you take these as an individual?Partner Network only, registered with a partner email. What the rule means, the three routes independents use, fees and discounts.7 min
- Registration to badgeFrom registration to Credly badgePartner Academy, Pearson VUE login, online or test centre, the ID name rule that forfeits fees, the 48-hour window, retakes, renewal.8 min
- SponsorshipHow to get sponsored with no partner employerCheck your own employer first, then a partner program you can apply to, and what a partner can see.3 min
- Exam factsCost, length, passing score60 items, 120 minutes, $125, and what a scaled 720 actually means.5 min
- BlueprintThe five domainsEvery domain, its published weight, and what the questions reward.6 min
- Exam formatHow the scenarios workFour scenarios drawn from a published bank of six, about 15 items each. What that means for how you sit it.7 min
- Straight answerAre there CCAR-F dumps?Why they backfire on a scenario-based exam, and what to use instead.6 min
- Straight answerExam dumps, in generalWhat dumps are, and five tests that tell one from legitimate practice questions.8 min
Practice inside your assistant
Claude and ChatGPT can quiz you on CCAR-F — without being able to give you the answer
All 60 free CCAR-F questions connect to the assistant you already use. It gets the question and not the answer key, so it cannot hand you the answer or be talked into agreeing with you. You commit first; then it explains why each option wins or loses.
Also here