Anthropic Academy: Production Skills the CCA-F Exam Tests

Anthropic launched the Claude Certified Architect, Foundations exam (CCA-F) on 12 March 2026, and the question we hear most from candidates is a version of the same thing: "I have finished the Academy courses. Am I ready?" The honest answer is: probably not yet. Academy course completion certificates and the CCA-F are different artefacts measuring different things. This post maps what the Academy teaches, where the exam goes further, and how to close the gap.

What does Anthropic Academy actually cover?

Anthropic Academy, delivered via Skilljar, offers a set of self-paced courses including Claude 101, the API fundamentals course, Claude Code, and MCP integration. Each course ends with a completion certificate. These certificates confirm you watched the material and passed the embedded knowledge checks; they do not appear on your Anthropic partner profile and carry no weight in the Claude Partner Network on their own.

The courses are genuinely useful. They introduce the Messages API request-response cycle, explain tool-use syntax, walk through Claude Code setup, and demonstrate basic MCP server configuration. For someone new to the Claude ecosystem, they are the right starting point.

What they do not do is test your ability to reason under pressure about production trade-offs. The CCA-F does exactly that: 60 scenario-based multiple-choice questions, each with one correct answer and three plausible distractors, scored on a 100-to-1000 scale with a passing mark of 720.

How does the CCA-F exam differ from course completion certificates?

The exam differs in three concrete ways.

First, format. Academy courses use low-stakes knowledge checks. The CCA-F is online-proctored or delivered at a test centre, costs $99 per attempt, and is time-limited. There is no "rewatch the video" option mid-question.

Second, depth. The exam blueprint covers five weighted domains:

Domain 1 alone accounts for more than a quarter of the exam. Academy courses introduce agentic concepts; the exam asks you to diagnose a broken multi-agent loop, choose between hub-and-spoke and flat orchestration, or decide when a coordinator should route dynamically versus follow a fixed pipeline. That is a different cognitive task.

Third, scoring. Anthropic does not publish the raw-to-scaled conversion, so we will not state an exact question count as the pass mark. What we can say is that 720 on a 100-to-1000 scale is a meaningful bar, and the exam consistently rewards deterministic solutions over probabilistic ones when stakes are high.

What prerequisites does the exam assume?

Anthropic's own guidance points to roughly six months of hands-on experience building agentic systems, or a comparable adjacent role such as solutions architecture or ML engineering with Claude API exposure. This is not a gatekeeping rule; there is no enforced prerequisite check at registration. It is a calibration signal.

If you have spent six months reading documentation but not shipping code, the exam will surface that gap quickly. The scenario questions assume you have encountered real failure modes: a subagent that loses attribution when synthesising across sources, a tool description that causes systematic misrouting, a session that degrades because stale context was never pruned. Academy courses describe these concepts; production experience makes the correct answer obvious.

For candidates who are earlier in their journey, our concept library at /concepts maps 174 atomic concepts to all five domains and 30 task statements, which is a structured way to build the mental models the exam tests.

Which Academy courses form the core study roadmap?

The four courses that map most directly to exam domains are:

1. Claude 101 covers model capabilities, the Messages API, and basic prompt construction. This underpins Domain 4 (Prompt Engineering & Structured Output) and parts of Domain 5 (Context Management & Reliability).
2. API fundamentals covers tool-use syntax, stop reason handling, and multi-turn conversation structure. This is foundational for Domain 1 and Domain 2.
3. Claude Code covers the three-level configuration hierarchy, CLAUDE.md authoring, and the Claude Code CLI. This maps directly to Domain 3.
4. MCP integration covers server configuration, scoping, environment variable expansion, and resource patterns. This maps to Domain 2.

The gap between "course complete" and "exam ready" is largest in Domain 1. Agentic Architecture & Orchestration at 27% of the exam covers concepts like parallel subagent spawning, coordinator dynamic subagent selection, and hook pipeline architecture that the Academy courses introduce but do not drill at exam depth.

How should you approach token economics and prompt caching for the exam?

Token economics questions appear primarily in Domain 5 (Context Management & Reliability, 15%) and at the edges of Domain 4. The exam does not ask you to recite pricing; it asks you to reason about trade-offs.

A representative scenario type: a production multi-agent system is making repeated calls with a large, stable system prompt. The question asks which technique reduces latency and cost most reliably. The correct answer is prompt caching, because it is deterministic and does not require restructuring the conversation. A probabilistic answer like "summarise the system prompt" introduces quality risk and is not the proportionate fix.

The exam's preference for deterministic, proportionate, root-cause solutions is consistent across domains. When you see a scenario where a probabilistic fix and a deterministic fix are both on offer, the deterministic one is almost always correct when stakes are high.

For context management specifically, the context management concepts section covers the stale context problem, summary injection patterns, and session management options in the detail the exam expects.

How do Claude Opus and Sonnet trade-offs appear on the exam?

Model selection questions appear in Domain 1 and Domain 4. The exam does not ask you to memorise benchmark scores. It asks you to articulate the trade-off in a given scenario.

The pattern is consistent: Opus offers higher reasoning capability at higher cost and latency; Sonnet offers a better cost-latency profile for high-volume or latency-sensitive tasks. The exam rewards candidates who can match model choice to task characteristics rather than defaulting to "use the most capable model."

A typical scenario: a coordinator agent needs to classify incoming requests into one of five categories before routing them to specialised subagents. The classification step runs on every request. Which model is appropriate? Sonnet, because the task is well-defined, high-volume, and does not require deep reasoning. Opus would be the right choice if the classification involved ambiguous, high-stakes decisions where errors are costly to recover from.

This reasoning pattern connects directly to model-driven vs pre-configured decision-making, a concept that appears across multiple Domain 1 task statements.

What does the exam test about MCP integration specifically?

Domain 2 (Tool Design & MCP Integration, 18%) tests both tool design principles and MCP-specific mechanics. The two are related but distinct.

On the tool design side, the exam focuses on:

• How tool descriptions function as the primary selection mechanism for the model
• When to split a broad tool into narrower, more specific tools
• How to diagnose systematic tool misrouting

On the MCP side, the exam focuses on:

• The scoping hierarchy (user, project, local)
• Environment variable expansion in MCP configuration files
• The isError flag pattern and how it differs from a valid empty result
• When to build a custom MCP server versus use an existing one

{
  "mcpServers": {
    "data-warehouse": {
      "command": "npx",
      "args": ["-y", "@company/mcp-warehouse"],
      "env": {
        "DW_API_KEY": "${DW_API_KEY}",
        "DW_ENDPOINT": "${DW_ENDPOINT}"
      }
    }
  }
}

The configuration above illustrates environment variable expansion in an MCP config. The exam tests whether you understand that hardcoding credentials here is a security failure, and that the correct pattern is to reference environment variables so that secrets stay out of version-controlled configuration files. See environment variable expansion in MCP config for the full treatment.

How does the exam validate production implementation skills?

As of 3 June 2026, more than 10,000 individuals have earned the CCA-F certification, and over 40,000 partner applicant firms have joined the Claude Partner Network. The volume signals that the market has accepted the certification as a meaningful credential, but the more important question for candidates is what the exam actually validates.

The CCA-F validates that you can reason correctly about production implementation decisions, not that you can recall transformer architecture theory. The scenario questions are grounded in real failure modes: an agentic loop that terminates prematurely because stop_reason was not inspected correctly, a multi-agent system where attribution is lost during synthesis, a Claude Code workflow where the wrong configuration scope causes a rule to be silently ignored.

This is why hands-on experience matters more than course completion for exam readiness. The distractors in CCA-F questions are designed to catch candidates who know the vocabulary but have not encountered the failure modes. If you have shipped a production multi-agent system, you will recognise the wrong answers immediately. If you have only read about them, the distractors will look plausible.

How do you register for the CCA-F exam?

Registration requires access to the Claude Partner Network. The path is:

1. Complete the Anthropic Academy courses via Skilljar (not a hard prerequisite, but the recommended preparation baseline).
2. Request Claude Partner Network access through the Anthropic partner portal.
3. Once approved, access the exam registration link through the Partner Network dashboard.
4. Pay the $99 exam fee. Tiered Anthropic partners receive a discounted first attempt.
5. Choose online-proctored delivery or a test centre, then schedule your sitting.

Anthropic has announced further architect, developer, and seller certifications planned for later in 2026, so the CCA-F is the first in a growing credential stack rather than a standalone artefact.

How should you use AI Skill Certs alongside the Academy?

AI Skill Certs (this platform) is independent of Anthropic. We are not affiliated with, endorsed by, or approved by Anthropic. What we offer is a structured way to close the gap between Academy course completion and exam readiness.

Our adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold, which means the platform keeps drilling a concept until you demonstrate reliable recall, not just one correct answer. Archie, our Socratic tutor, guides you through the reasoning behind each concept with graduated hints rather than giving you the answer directly.

Practice exams are 60 questions, scored 100 to 1000 with 720 as the passing bar, matching the real exam format exactly. The concept library covers 174 atomic concepts mapped to all five domains and 30 task statements, so you can identify and address weak areas systematically rather than re-reading course material you already know.

The most efficient preparation path we have observed: complete the Academy courses first to build vocabulary, then use the practice exams to identify domain gaps, then drill the specific concepts where your score is below threshold.