Pass the Claude Architect Foundations exam
A primer on what this exam really is, then five guides — one per exam domain, covering every subdomain of the official exam guide — the architecture concepts taught plainly, the flags and files the exam quotes, and the traps it sets. No pass guarantee; just nothing missing. About 81 minutes of reading, and the credential lasts a year.
The exam at a glance
Claude Certified Architect — Foundations (CCAR-F) is the entry point to Anthropic’s architect track — and the hardest of the three certifications despite the name. It tests foundational knowledge across four surfaces — Claude Code, the Claude Agent SDK, the Claude API and MCP — and a large share of its questions are concrete and technical: CLI flags, parameter values, file locations. We sat all three; this is the one that deserves the most respect.
If you don’t pass: wait 14 days before your second attempt, 30 days before your third, 90 days before your fourth — up to four attempts per rolling year. The full booking timeline and the exam-day tactics are in the prep-plan post; these guides are the study material that post promised.
The five domains
Study where the marks are. Agentic architecture and orchestration (27%), Claude Code configuration and workflows (20%) and Prompt engineering and structured output (20%) — the three highest-weighted domains — carry 67% of the exam between them, so read them first. Note the domain order below follows the official exam guide, not the weights: tool design is Domain 02 even though two later domains outweigh it. Context management and reliability is the lightest at 15% — a study session, not a glance. Each guide page lists the modules it covers, in exam-guide order.
What the architect exam actually is (harder than its name), how it differs from the developer one, and the four surfaces it tests — the floor the five guides build on. Five minutes; skip it if you have shipped with the Agent SDK.
| Domain | Weight | Modules | Read | Study priority |
|---|---|---|---|---|
| 01Agentic architecture and orchestration | 27% | 7 | 19 min | 1st of 5 |
| 02Tool design and MCP integration | 18% | 5 | 12 min | 4th of 5 |
| 03Claude Code configuration and workflows | 20% | 6 | 12 min | 2nd of 5 |
| 04Prompt engineering and structured output | 20% | 6 | 12 min | 3rd of 5 |
| 05Context management and reliability | 15% | 6 | 13 min | 5th of 5 |
The heart of the exam: who decides the next step, the loop an agent actually runs, and how coordinator–subagent systems pass context, enforce rules and survive failures.
Tools are the agent’s hands: how to write the contract (name, description, schema), how to fail usefully, how to distribute tools across agents, and how MCP servers plug in.
Which file outranks which: CLAUDE.md for guidance, settings.json for enforcement, slash commands and skills, plan mode, checkpoints, and unattended headless runs.
Getting reliable, repeatable behaviour: explicit acceptance criteria, few-shot examples that teach, schemas that enforce, retries that teach back, batching that halves cost, and review that runs twice.
The window is finite: compaction, context editing and the memory tool; escalation ladders; how errors travel between agents; exploring big codebases; calibration and provenance.
How to use this guide
Read the domains in weight order — that is what the study-order toggle above is for — because hours spent should follow marks available. Each page stands alone: the concepts taught from the ground up, the specific names and numbers the exam quotes, and a traps section collecting the wrong answers that sound sensible.
Prerequisites: the developer-track floor. If Claude itself — or APIs in general — is new to you, start with the CCDV-F primer first: this guide assumes you know what an API call looks like and that a token is. If you build with Claude, the Claude Agent SDK or Claude Code already, you have the floor — but budget real study anyway: daily use hides exactly the file paths and defaults this exam tests.
How this was written and checked: by Fabs, who sat all three Anthropic certifications in 2026 — Developer, Architect Foundations, Architect Professional — and wrote up each one. The structure follows the official exam guide’s five domains and every subdomain in them; every fact is anchored in Anthropic’s documentation as it stood in September–October 2026; and coverage was cross-checked against the independent CertSafari and dnacenta guides. Exactly what is confirmed by Anthropic, what a real sitting confirmed, and what is our judgment alone is sorted claim by claim on the evidence and transparency page, alongside the six official scenario contexts. We are not affiliated with or endorsed by Anthropic.
Model names, prices and policies move — if you are reading this in six months, re-check anything numeric against the docs before you memorise it.
One honest caveat: a certificate proves you studied, not that you can design. But the studying here is the real thing — the same architecture decisions we make building agent systems for clients — and if you pass, you get a year of arguing about agents with a badge behind your name.
Where to verify us
Anthropic’s own material is free and it is the source of truth. If you have the hours, start there; this guide is the compression — same syllabus, about 81 minutes, the numbers in one place, the traps marked. Every claim on this page is checkable against the links below. Already holding the developer certification? The CCDV-F guide covers that exam the same way.
Registration, the exam guide (the source of truth for domains and weights), and policies.
The free official courses. Weight your hours to the big domains: the agent courses, Claude Code, and MCP.
Where every fact in this guide is ultimately anchored. Read it like it’s 2003 and you have an exam tomorrow.
480 free questions with explanations, keyed to the exam guide’s subdomains. Useful as a second pass.
A free unofficial GitHub guide covering all five domains with a study plan. We cross-checked ours against it.