Claude Certified Architect (CCA) Exam Guide: All Five Domains on One Diagram

July 24, 2026 · Claude Certified Architect (CCA) Prep (part 1)

▶ Watch on YouTube & subscribe to The Stack Underflow

You open the Claude Certified Architect exam expecting a definitions quiz — what is an agent, what is MCP, what is a system prompt — and instead you get a broken production scenario: an agent calls a tool, ignores the result, and answers from memory anyway. What would you change? That’s the whole exam. Sixty scenarios like that, drawn from four of six possible situations, two hours, no docs, no Claude, no browser. It is not testing whether you memorized a term. It is testing whether you can look at a misbehaving agent and know which of five levers to pull.

The one-sentence version: the Claude Certified Architect exam has five domains, and every single scenario question is really asking you to design, equip, operate, instruct, or sustain a Claude agent — learn that map and the exam becomes a reading test, not a memory test.

The Architect’s Map

Picture a single Claude agent at the center, and five things you can do to it wrapped around it like sectors on a dial. Each sector is sized by how much of the exam it accounts for — so before you study a single flashcard, you already know where to spend your time.

                        ┌─────────────────────────────────────────────┐
                        │            THE ARCHITECT'S MAP               │
                        │        (everything = 1 of 5 things)          │
                        └─────────────────────────────────────────────┘

        D1 · AGENTIC ARCHITECTURE & ORCHESTRATION .. 27%   "DESIGN it"   ▓▓▓▓▓▓▓▓▓ (biggest)
        D3 · CLAUDE CODE CONFIG & WORKFLOWS ........ 20%   "OPERATE it"  ▓▓▓▓▓▓▓
        D4 · PROMPT ENGINEERING & STRUCTURED OUTPUT  20%   "INSTRUCT it" ▓▓▓▓▓▓▓
        D2 · TOOL DESIGN & MCP INTEGRATION ......... 18%   "EQUIP it"    ▓▓▓▓▓▓
        D5 · CONTEXT MANAGEMENT & RELIABILITY ...... 15%   "SUSTAIN it"  ▓▓▓▓▓
                                                     ────
                                                     100%  → one Claude agent that works
   (exam = 4 scenarios drawn at random from a pool of 6 · 60 Q · 120 min · pass 720 · Pearson VUE · $125 · retakes available)

Every option on every multiple-choice question is a plausible-sounding fix. The exam isn’t rewarding the person who read the most documentation — it’s rewarding the person who can identify which of these five sectors the broken scenario actually lives in, because that’s the only way to tell a merely-plausible answer from the correct one.

Domain 1 — Agentic Architecture & Orchestration (27%, “Design it”)

The single heaviest domain, and it’s heavy for a reason: it’s the shape of the agent itself. This is where you’re tested on the agent loop — the observe-act-observe cycle, and specifically what happens when a step in that loop is skipped (an agent that calls a tool and ignores the result, for instance, is a D1 failure, not a D2 or D4 one). It covers coordinator-subagent patterns for when two agents disagree on an answer, task decomposition for when a single agent’s context window can’t hold a whole job, session state versus long-term memory, and hooks used as workflow enforcement rather than as a convenience. It also covers the judgment call of when not to build an agent loop at all — plenty of scenario questions are really testing whether you’ll reach for a deterministic workflow instead of over-engineering an agentic one.

Because it’s nearly a third of the exam, D1 is the correct place to start studying, not an afterthought.

Domain 2 — Tool Design & MCP Integration (18%, “Equip it”)

D2 is how the agent reaches the outside world. It covers tool interface design — writing tool descriptions specific enough that Claude picks the right one among dozens, typing parameters so “next Tuesday” doesn’t get passed as a raw string where a date is expected, and deciding what’s required versus optional. It covers structured error contracts (what a tool returns when it fails, not just when it succeeds), pagination for tools that could return five thousand rows, and the trade-off at the center of this domain: when a tool interface should live inline versus behind an MCP server, and how MCP authentication works across multiple tenants. If Domain 1 is about the agent’s shape, Domain 2 is about its hands.

Domain 3 — Claude Code Configuration & Workflows (20%, “Operate it”)

D3 is the only domain that’s entirely about the tooling layer rather than the model layer — how you actually run and govern an agent day to day inside Claude Code. This is the CLAUDE.md hierarchy (project-level conventions versus a rule that should stay local to you, not your whole team), the difference between an instruction and a hook (only one of them runs unconditionally, every time), slash commands versus Skills, the permissions allowlist and how enterprise-level denies interact with project-level allows, what belongs in settings.local.json and why it should never be committed, headless mode for running Claude Code in CI with no human present, and scoped custom subagents built for one narrow job. Anyone who’s used Claude Code daily has a real advantage here — this domain rewards operational familiarity more than any other.

Domain 4 — Prompt Engineering & Structured Output (20%, “Instruct it”)

D4 is how you get reliable, machine-usable behavior out of the model rather than a plausible-sounding paragraph. It covers forcing valid JSON output and validating the types that come back (a number arriving as a string is a real, recurring failure mode), keeping chain-of-thought reasoning available to the model while excluding it from the final structured output, building genuinely diverse few-shot examples instead of five near-duplicates, choosing between a forced and an auto tool_choice, designing bounded retry loops so a validation failure doesn’t spin forever, batch processing for large overnight jobs, where in the prompt an instruction needs to sit to actually get followed, and preferring an explicit “unknown” field over a hallucinated one when the output feeds a downstream program rather than a person.

Domain 5 — Context Management & Reliability (15%, “Sustain it”)

The smallest slice by weight, but the one that decides whether a system that works in a demo keeps working in a long-running production task. D5 covers lost-in-the-middle effects and context rot, chunking and decomposition when a task overflows the context window, prompt caching (and the prefix rules that determine why a cache keeps missing), grounding and citations so a claim can be traced back to a source, error propagation — an agent that fails a step and then answers confidently anyway is a D5 failure — escalation patterns for handing a task to a human, confidence calibration, persisting an agent’s scratchpad across a context compaction event, and budgeting the context window between conversation history and the actual task.

How the exam maps a scenario to a domain

The table below is the fast version of the whole map — read a scenario, ask “what’s actually broken,” and match it to a row.

Notice that this table is really the exam’s own reframe: it’s never asking “do you know what MCP is.” It’s asking “given this specific broken behavior, which lever fixes it, and why not the other four.”

Exam logistics (verify before you book)

As of mid-2026, the format looks like this — but exam logistics move faster than exam content, so treat every number here as something to re-confirm on Anthropic’s official pages before you register.

• Delivery: Pearson VUE, in a testing center — no docs, no Claude, no browser in the room.
• Format: 60 scenario questions, drawn as 4 scenarios from a pool of 6, in 120 minutes.
• Scoring: a scaled score from 100–1000; passing is 720. That is not the same as answering 72% of questions correctly — it’s a scaled score, not a raw percentage, so don’t try to back-calculate a “questions I can afford to miss” number from it.
• Cost and retakes: $125 per attempt, with retakes available (this reportedly changed from a one-attempt, $99 policy in mid-2026 — another reason to verify current terms directly).

None of these numbers are the point of studying. They’re logistics. The point is the map above: five things you do to a Claude agent, weighted by how often the exam asks about each one.

A study order that matches the weights

Given the weighting, the efficient order is D1 first (27%, the single biggest lever), then D3 and D4 together (20% each — operating the tool and instructing the model are both “how you actually get work done day to day”), then D2 (18%, equipping the agent with tools), and D5 last by weight but not by importance (15% — the domain that decides whether your system survives contact with a long real task). Don’t skip D5 because it’s smallest; a system that fails D5 concerns fails silently, in production, days after launch.

Don’t memorize the answer to any single practice question. Understand which sector of the map it lives in, and why the other four sectors don’t fit — that’s the skill the exam is actually measuring, and it’s the skill that transfers to building real agents once the exam is behind you.

Not affiliated with, authorized, or endorsed by Anthropic. “Claude” and “Claude Certified Architect” are trademarks of Anthropic. These are original practice questions for study, not real exam content — verify current exam details on Anthropic’s official pages.

Where this fits in the series

This is the keystone of the Claude Certified Architect Prep series — every other episode zooms into one sector of the map above. If you’re weak on how tools reach the outside world, start with MCP Explained on One Diagram: The USB-C Port for AI Tools, a D2 concept that shows up constantly disguised as other domains’ questions. If you’re unsure how retrieval and context strategy relate, RAG vs Context Engineering: Why ‘RAG Is Dead’ and What Replaces It untangles a distinction the exam leans on across both D2 and D5. And if D1 is genuinely your weakest sector — which, given it’s 27% of the exam, it’s worth checking — start with The Agent Loop Explained: Why Ignoring Tool Results Breaks Your Agent, the exact failure mode from the cold-open scenario above. Browse all tutorials for the full per-domain practice set.