Anthropic Learning Path: CCA-F Architect Exam Roadmap

The anthropic learning path for the Claude Certified Architect, Foundations exam (CCAR-F) is not a single course. It is a structured sequence of five domains, each demanding practical judgment rather than recall, spread across 60 scenario-based items and a 120-minute sitting. This guide maps that sequence, weights each domain honestly, and tells you where to spend your study hours.

We are an independent prep platform. AI Skill Certs is not affiliated with, endorsed by, or approved by Anthropic.

What is the CCAR-F and where does it sit in the Claude Partner Network?

The Claude Certified Architect, Foundations (CCAR-F) is the architect track of Anthropic's Foundations certification programme, launched 12 March 2026. It costs $125 USD per attempt, runs for 120 minutes, and contains 60 items. The passing score is 720 on a 100-to-1000 scale. Per Anthropic's exam guide, every item is scenario-based and tests practical judgment, not recall.

The CCAR-F sits inside the Claude Partner Network, a $100 million programme that had attracted more than 40,000 partner applicant firms and 10,000+ certified individuals as of 3 June 2026. Three other proctored Pearson VUE tracks run alongside it:

The credential is valid for 12 months from the date it is awarded. Tiered Claude Partner Network partners receive a discounted first attempt.

How is the exam structured across its five domains?

The CCAR-F covers five domains and 30 task statements. Domain weights are published by Anthropic and drive every study-hour allocation decision you should make:

Domains 1, 3, and 4 together account for 67% of the exam. If your study time is constrained, those three domains return the highest yield per hour. Domain 5, at 15%, is the smallest slice but contains reliability and escalation patterns that appear as the deciding factor in many scenario items.

Each sitting draws four scenarios at random from a bank of six, so you cannot predict which scenario cluster will appear. Breadth across all five domains is not optional.

What does the anthropic learning path look like in practice?

A defensible anthropic learning path moves from foundational mechanics to applied architecture, then to reliability and edge cases. We recommend four phases.

Phase 1: Agentic loop mechanics (Domain 1, weeks 1 to 2)

Domain 1 carries 27% of the exam weight, making it the single largest investment. Start here. The core skill is understanding how the Messages API request-response cycle drives agentic behaviour, how to inspect the stop_reason field to decide whether to continue a loop, and how to append tool results correctly before the next turn.

From there, move to orchestration patterns. The hub-and-spoke architecture is the dominant multi-agent topology on the exam: a coordinator receives a task, selects subagents dynamically, and synthesises results. Understand coordinator responsibilities precisely, including when the coordinator should route errors versus escalate to a human.

The exam consistently rewards deterministic solutions over probabilistic ones when stakes are high. That principle surfaces most visibly in Domain 1 scenarios involving prerequisite gate design and workflow enforcement.

Phase 2: Tool design and Claude Code configuration (Domains 2 and 3, weeks 3 to 4)

Domain 2 (18%) and Domain 3 (20%) share a common theme: configuration precision determines runtime behaviour. In Domain 2, the central skill is writing tool descriptions that function as selection mechanisms. When Claude misroutes a tool call, the first diagnostic question is whether the description is ambiguous, not whether the model is broken. Study tool splitting for specificity and the MCP isError flag pattern for structured error responses.

Domain 3 focuses on Claude Code configuration. The three-level configuration hierarchy (global, project, path-scoped) is the conceptual anchor. Exam items test whether you can diagnose a misconfiguration by identifying which level is overriding which, and whether a given rule belongs in a CLAUDE.md file or a custom skill with YAML front matter. Version control implications of each level are a recurring scenario type.

Phase 3: Prompt engineering and structured output (Domain 4, weeks 5 to 6)

Domain 4 (20%) tests your ability to design prompts that produce reliable, parseable output at scale. The exam rewards few-shot examples as the highest-leverage technique for ambiguous extraction tasks, and it tests JSON schema design for preventing fabrication in structured pipelines.

A common scenario type: a batch extraction pipeline is returning inconsistent field values. The correct fix is almost never "increase temperature" or "add more instructions." It is to add a well-chosen few-shot example that demonstrates the boundary case, or to tighten the schema to eliminate the ambiguous field. Study prompt engineering concepts with that diagnostic lens.

The attention dilution problem also appears in Domain 4 scenarios: when a prompt grows too long, model attention spreads across irrelevant context and output quality degrades. The fix is decomposition, not compression.

Phase 4: Context management and reliability (Domain 5, weeks 7 to 8)

Domain 5 (15%) is the smallest domain but the one most likely to separate passing from failing candidates who have studied only the happy path. The core scenarios involve stale context, session management decisions, and escalation routing.

The stale context problem arises when a long-running session accumulates outdated facts that contradict later tool results. The exam tests whether you can identify this as the root cause and apply summary injection for fresh sessions rather than simply retrying the same call.

Escalation design is the other major topic: when should an agent resolve autonomously versus hand off to a human? The exam rewards confidence-based routing with explicit thresholds, not vague "when uncertain, ask" instructions. Structured handoff to human agents is a concrete pattern worth studying in full.

How should you allocate study hours across the path?

A candidate with 60 total study hours and no prior agentic architecture experience should allocate roughly as follows:

These are starting allocations, not fixed rules. After a diagnostic practice exam, shift hours toward domains where your percent-correct score is below 70%. The score report from each practice sitting gives you percent-correct by domain, which is the same breakdown the real exam report provides.

What exam mechanics should you know before sitting?

Several mechanics affect strategy in ways that are easy to overlook.

Scenario bank. Each sitting draws four scenarios at random from a bank of six. You cannot predict which four appear. Preparing for only the scenarios you find interesting is a risk you should not take.

Multiple-response items. Some items ask you to select two or three correct answers. Each item states how many to select. Partial credit is not confirmed by Anthropic, so treat each item as requiring all correct selections.

Raw-to-scaled conversion. Anthropic does not publish the raw-to-scaled conversion. The passing score is 720 on the 100-to-1000 scale. On a linear read this is roughly 41 to 42 of 60 questions, but we do not state that as the pass mark because the conversion is not linear and is not published.

Delivery. The exam is available online-proctored or at a Pearson VUE test centre.

How does AI Skill Certs support this learning path?

Our concept library at /concepts covers 174 atomic concepts mapped to all five CCAR-F domains and all 30 task statements. Each concept page is a focused, scenario-grounded explanation, not a summary of documentation.

The adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold. When you answer a question, the engine updates its estimate of your mastery for every concept that question touches, then surfaces the concept where your estimated mastery is lowest. You spend time on gaps, not on material you already know.

Archie, our Socratic tutor, never gives the answer directly. It guides with graduated hints so you build the reasoning pattern rather than memorising a response. That matters for a scenario-based exam where the surface details change but the underlying judgment pattern repeats.

Practice exams mirror the real format: 60 questions, scored 100 to 1000, with 720 as the passing bar and a percent-correct breakdown by domain after each sitting.

What is the right order to tackle the concept library?

We recommend following domain weight order for the first pass through the concept library, then using your practice exam domain scores to direct the second pass.

For Domain 1, start with the Messages API request-response cycle, then stop_reason field inspection, then tool result appending. These three concepts underpin every agentic loop scenario on the exam. From there, move to orchestration: hub-and-spoke architecture, parallel subagent spawning, and multi-agent error handling and routing.

For Domain 2, the tool descriptions as selection mechanism concept is the entry point. Everything else in Domain 2 is a variation on that principle applied to different failure modes.

For Domain 5, the session management options concept and the when to resume vs fork vs fresh start decision framework are the two most exam-relevant pages.

What common mistakes do candidates make on this path?

Studying documentation, not judgment. The exam does not test whether you can recite API parameters. It tests whether you can diagnose a broken system and choose a proportionate fix. Every study session should end with a question: "Given this scenario, what is the root cause, and what is the minimum change that resolves it?"

Ignoring Domain 5. At 15%, it feels skippable. In practice, context management and escalation scenarios are the ones where candidates who have only studied the happy path lose points they cannot recover elsewhere.

Treating the scenario bank as predictable. Four of six scenarios appear per sitting. Candidates who prepare deeply for two or three scenario types and skim the rest are gambling on the draw.

Conflating the CCAR-F with the CCAO-F. The Associate exam (CCAO-F) costs $99 and is a separate track. The Architect exam costs $125. These are different credentials with different domain structures.

Stopping after one practice exam. The percent-correct by domain on your first practice sitting is diagnostic data. The second and third sittings, targeted at your weakest domains, are where the score improvement happens.