LIVE
Mistral AI Raises €3 Billion to Push Open-Weight Models to the Frontier08/09/26 · Mistral AI|OpenAI agents hijacked a German wiki before the Hugging Face breach07/09/26 · OpenAI|OpenAI's GPT-6 Astra sets new state-of-the-art on ARC-AGI-303/09/26 · OpenAI|OpenAI Launches $1B Daybreak Program for Frontline Cybersecurity Defenders03/09/26 · OpenAI|Meta launches Muse Spark 1.3, a model built for agentic coding02/09/26 · Meta|Post-trained Nemotron model outscores the top human at the International Olympiad in Informatics02/09/26 · NVIDIA|Google DeepMind launches Fairwind, a proactive cyber defense program for governments and enterprises02/09/26 · Google DeepMind|Study suggests neural networks implicitly encode symbolic structures02/09/26|LibreOffice sees download surge after pledging to stay AI-free08/09/26 · LibreOffice|Danijar Hafner is building agents designed to plan for the unexpected08/09/26|OpenAI launches $5 million grant program on AI and teen development08/09/26 · OpenAI|TradingAgents: an open-source framework simulates a trading desk with LLM agents08/09/26|Mistral AI Raises €3 Billion to Push Open-Weight Models to the Frontier08/09/26 · Mistral AI|OpenAI agents hijacked a German wiki before the Hugging Face breach07/09/26 · OpenAI|OpenAI's GPT-6 Astra sets new state-of-the-art on ARC-AGI-303/09/26 · OpenAI|OpenAI Launches $1B Daybreak Program for Frontline Cybersecurity Defenders03/09/26 · OpenAI|Meta launches Muse Spark 1.3, a model built for agentic coding02/09/26 · Meta|Post-trained Nemotron model outscores the top human at the International Olympiad in Informatics02/09/26 · NVIDIA|Google DeepMind launches Fairwind, a proactive cyber defense program for governments and enterprises02/09/26 · Google DeepMind|Study suggests neural networks implicitly encode symbolic structures02/09/26|LibreOffice sees download surge after pledging to stay AI-free08/09/26 · LibreOffice|Danijar Hafner is building agents designed to plan for the unexpected08/09/26|OpenAI launches $5 million grant program on AI and teen development08/09/26 · OpenAI|TradingAgents: an open-source framework simulates a trading desk with LLM agents08/09/26|
+ nAIvigateA SHEET FOR PEOPLE WITHOUT 2 HOURS
+ nAIvigate · fiches / cca-f-5-domaines-une-page
▣ fiche-006.md ×
□ index.md
ContentsVerdictThe 5 domainsThe 10 reflexesThe recurring distractorsThe numbers to rememberD-7 plan

CCA-F: 15 days of revision? One page first.

60 questions, 120 minutes, 720 out of 1,000, $125. The exam doesn't test whether you can build an agent: it tests whether you choose the right mechanism for a production scenario — and nine times out of ten, the trap is a sentence in the prompt where it expects code. This sheet condenses our five domain-by-domain lessons (≈38,000 words) into what you need to keep in mind on exam day.

Verdict

the specs in 10 seconds5 rows
FORMAT60 scenario-based questions, multiple choice and multiple response (the prompt says how many to tick). 120 min, English.CCAR-F
PASS720 / 1000, weighted by domain. No penalty: never leave a blank.~72%
COST$125 per attempt ($99 at launch). Retakes: +14, +30, +90 days · max 4 per 12 months.PEARSON VUE
VALIDITY12 months, free non-proctored renewal before expiry. Credly badge.12 MONTHS
SCENARIOS4 drawn from 6: customer support · code gen · multi-agent research · dev productivity · CI/CD · document extraction.4 / 6

The 5 domains

revise by weight, not in order5 rows
D1Agentic Architecture — stop_reason, hub-and-spoke, Task, gates & hooks, chaining vs dynamic, --resume / fork.27%
D3Claude Code — CLAUDE.md × 3 levels, .claude/rules/ + paths, commands vs skills, plan mode, -p / --output-format json.20%
D4Prompt & Structured Output — criteria vs adjectives, tool_use + schema, nullable, retry with feedback, Batch -50% / 24h.20%
D2Tool Design & MCP — the description picks the tool, structured errors, 4-5 tools/role, tool_choice, .mcp.json vs ~/.claude.json.18%
D5Context & Reliability — window budget, structured summary, partial with gaps, enriched error, constrained source_ids.15%

The 10 reflexes

what the exam expects, every time10 rows
guaranteeReal consequence (money, data, compliance) → gate or hook, never a sentence in the prompt.D1
end_turnThe loop continues on stop_reason == tool_use, stops on end_turn. No counter, no "DONE".D1
coordinatorIncomplete coverage, correct subagents → decomposition too narrow. Everything through the hub.D1
descriptionWrong tool called → enrich the description (formats, examples, boundaries). No router, no few-shot.D2
isRetryableTool error = category + isRetryable + message. Legitimate empty ≠ error.D2
projectShared, clone, new member → project level (CLAUDE.md, .claude/, .mcp.json). Under ~/ = personal.D3
globFiles spread by type → .claude/rules/ + paths, not one CLAUDE.md per folder.D3
nullableInvented value for an absent field → nullable schema + "null if absent". Not an instruction.D4
independenceReviewing its own code → independent instance, never the session that generated.D3/D4
partialA source fails → partial with explicit gaps. Never "nothing", never "complete".D5

The recurring distractors

if you see this in an option, eliminate7 rows
"strengthen the prompt"Probabilistic. The exam wants deterministic when it matters.
"bigger model"Pushes the wall back, not the slope. The problem is almost always the instruction or the management.
"cap at N turns"Safety net, not a stop condition.
"read everything first"Saturates the window before the first decision. Grep → Read, incremental.
"--batch", CLAUDE_HEADLESSDon't exist. The flag is -p. Invented features are always wrong.
"majority vote"Suppresses real intermittent findings. One instance per perspective + an aggregator.
"return [] as success"Silent suppression. The worst anti-pattern of Domain 5.
720OUT OF 1000 · 60 QUESTIONS · 120 MIN

Scoring is weighted: failing Domain 1 (27%) costs twice as much as failing Domain 5 (15%). Revise in weight order: D1, then D3 and D4, then D2, then D5.

The numbers to remember

the ones distractors play with7 rows
-50% · 24hBatch API: discount and turnaround. custom_id to match results (order not guaranteed).
4-5Tools per role. At 18, selection collapses.
3CLAUDE.md levels: user (~/), project (root, version-controlled), directory.
3SKILL.md frontmatter options: context: fork · allowed-tools · argument-hint.
3tool_choice modes: auto (text possible) · any (a tool) · forced (this tool).
4Error categories: transient · validation · business · permission.
12 / 3012 months of credential validity · 30 days = the typical business rule of the support scenario (refund past deadline → business error, no retry).

D-7 plan

one week, five evenings7 rows
D-7Official exam guide v1.0 in full + the 6 scenarios. It's a course on its own.2 H
D-6Lesson D1 + its 5 corrected questions. The domain that fails people.1 H 30
D-5Lesson D3 and D4 — the most factual, the most profitable.2 H
D-4Lesson D2 and D5 — the night-before checklists suffice if D1 is solid.1 H 30
D-3Anthropic Academy mock exam under conditions: 60 questions, 120 min, timed.2 H
D-2Re-read every mock mistake → identify the distractor you picked and the reflex you missed.1 H
D-1The 5 night-before checklists, in weight order. Nothing new. Sleep.45 MIN

Go deeper

fundamentals, in Learn6 links
Certifying an entire team? nAIvigate Studio prepares your architects across all five domains and delivers, in a Sprint, the real Claude system to go with it — gates, hooks, schemas, observability.
TALK TO nAIvigate STUDIO →
⎇ mainfiche #006 · certificationsources: Guide d'examen officiel CCAR-F v1.0 (juil. 2026) · Pearson VUE · Anthropic Academy · vérifié 04.09.2026
Follow nAIvigate | naivigate.com/en/fiches