Claude Code Agents: The CCAR-F Architect Exam Guide

What are Claude Code agents, and why do they dominate the CCAR-F exam?

Claude Code agents are autonomous or semi-autonomous processes that use Claude as their reasoning engine, invoke tools in a loop, and act on real systems rather than simply returning text. On the Claude Certified Architect, Foundations exam (CCAR-F, $125, 60 items, 120-minute limit), they are not a single topic but the connective tissue across all five domains. Domain 1 alone, Agentic Architecture & Orchestration, carries 27% of the exam weight, and agent-adjacent skills appear in every other domain too.

Understanding how claude code agents work at the implementation level, not just the conceptual level, is what separates candidates who score above the 720 passing mark from those who fall short.

How does the CCAR-F exam weight agent-related domains?

The exam draws from five domains with fixed percentage weights. The table below shows each domain, its weight, and the primary agent skill it tests.

Source: Anthropic CCAR-F exam guide, 12 March 2026.

Domain 1 and Domain 3 together account for 47% of the exam. If you are short on study time, those two domains return the highest yield per hour invested.

What is the agentic loop, and how does the exam test it?

The agentic loop is the repeating cycle in which Claude receives a message, decides whether to call a tool or return a final answer, receives the tool result, and continues until the task is complete or a stopping condition fires. The exam tests this loop at the level of practical diagnosis, not theory.

A typical scenario presents a loop that terminates too early or runs indefinitely and asks you to identify the root cause. The two most common failure modes are:

1. Premature termination caused by a stop_reason of end_turn when the model should have called another tool. This usually traces to a missing or ambiguous tool description, or a system prompt that implies the task is complete before it is. See our concept on Debugging Premature Loop Termination for the diagnostic checklist.

2. Infinite loops caused by a tool that always returns an error the model cannot resolve, combined with no maximum-iteration guard. The exam consistently rewards adding a deterministic iteration cap over probabilistic retry logic when stakes are high.

The stop_reason Field Inspection concept covers the exact field values (tool_use, end_turn, max_tokens, stop_sequence) and what each implies for loop control.

How do multi-agent topologies appear on the exam?

The CCAR-F exam uses scenario-based items, and multi-agent scenarios are the most complex. Each sitting draws four scenarios at random from a bank of six, so you will almost certainly face at least one multi-agent orchestration scenario.

The dominant topology tested is hub-and-spoke: a coordinator agent receives the user request, decomposes it, delegates subtasks to specialised subagents, and synthesises results. The exam tests three specific failure modes within this topology:

• Narrow decomposition failure: the coordinator assigns subtasks so tightly scoped that no subagent has enough context to handle edge cases. Our Narrow Decomposition Failure concept explains the diagnostic pattern.
• Attribution loss in synthesis: when the coordinator merges subagent outputs, source provenance is dropped, making the final answer unverifiable.
• Subagent context isolation: each subagent receives only the context it needs, not the full conversation. Violating this principle causes token waste and, in high-stakes workflows, security leakage.

The Hub-and-Spoke Architecture concept maps each of these failure modes to the task statements in Domain 1.

What does Domain 3 test about Claude Code configuration?

Domain 3 (Claude Code Configuration & Workflows, 20%) tests the three-level configuration hierarchy: project-level CLAUDE.md, user-level settings, and enterprise policy. The exam asks you to identify which level controls a given behaviour and what happens when levels conflict.

The most frequently tested sub-topics are:

• Hooks: PreToolUse and PostToolUse hooks intercept tool calls programmatically. The exam distinguishes hooks from prompt-based instructions: hooks are deterministic and cannot be overridden by the model; prompt instructions are probabilistic and can be ignored under adversarial or ambiguous conditions. When the scenario involves compliance or audit requirements, the correct answer is almost always a hook, not a prompt instruction.

• Slash commands and skills: custom slash commands let teams encode repeatable workflows. The exam tests when a skill is the right abstraction versus when a plain CLAUDE.md instruction suffices.

• Version control implications: CLAUDE.md files committed to a repository apply to every contributor. The exam tests the security and consistency implications of this, particularly in CI/CD pipelines.

A fenced example of a minimal PostToolUse hook that normalises tool output before it re-enters the context:

# hooks/post_tool_use.py
import json, sys

def normalise(result: dict) -> dict:
    # Strip internal metadata fields before Claude sees the result
    return {k: v for k, v in result.items() if not k.startswith("_")}

if __name__ == "__main__":
    raw = json.load(sys.stdin)
    print(json.dumps(normalise(raw)))

The Three-Level Configuration Hierarchy concept covers the precedence rules in full.

How does tool design affect agent reliability?

Tool design (Domain 2, 18%) is where many candidates lose points they did not expect to lose. The exam does not ask you to write tool implementations; it asks you to diagnose why an agent is calling the wrong tool or mishandling an error.

The two highest-yield concepts are:

Tool descriptions as the selection mechanism. Claude selects tools based on their descriptions, not their names. A vague description causes misrouting. The fix is almost always a description rewrite, not a system prompt addition, because description changes are lower-effort and higher-leverage. The Tool Descriptions as Selection Mechanism concept has the exact rewrite pattern.

MCP error propagation. When an MCP tool fails, the isError flag must be set to true in the response. If it is not, Claude treats the error payload as a valid result and continues reasoning on bad data. The exam tests four error categories and the correct structured metadata for each.

{
  "content": [
    {
      "type": "text",
      "text": "Database connection timed out after 5000ms."
    }
  ],
  "isError": true
}

Returning isError: false with an error message in the content field is the most common wrong answer in Domain 2 scenarios.

What context management patterns does the exam test?

Domain 5 (Context Management & Reliability, 15%) tests what happens to agent performance as context windows fill. The core problem is the attention dilution problem: as the context grows, Claude's effective attention to early instructions weakens. The exam tests three mitigations:

The exam rewards proportionate fixes. If the scenario describes a context that is 80% full with mostly irrelevant tool results, the correct answer is to trim tool results before they re-enter the context, not to start a fresh session and lose all accumulated state.

How should you approach exam scenarios involving agent safety?

The CCAR-F exam consistently applies three principles when agent safety is at stake:

1. Deterministic over probabilistic. When a constraint must hold without exception (compliance, data deletion, financial authorisation), implement it in a hook or a programmatic gate, not a prompt instruction.
2. Proportionate fixes. Match the intervention to the severity of the failure. A tool that occasionally returns a stale result needs a cache-invalidation fix, not a full architecture redesign.
3. Root-cause tracing. Scenarios describe symptoms. The correct answer identifies the root cause, not the most visible symptom. A loop that terminates early is a symptom; a missing tool description is a root cause.

These principles appear across all five domains and are the clearest signal of what the exam rewards. When two answer choices both fix the symptom, the one that addresses the root cause is correct.

How does our platform help you prepare for Claude Code agent scenarios?

AI Skill Certs is an independent adaptive prep platform (not affiliated with or endorsed by Anthropic). Our CCAR-F preparation includes:

• 174 atomic concepts mapped to all five domains and 30 task statements, available at /concepts. Each concept is a standalone, testable unit of knowledge.
• Practice exams that mirror the real format: 60 scenario-based items, scored 100 to 1000 with 720 as the passing bar.
• Archie, our Socratic tutor, which guides you through agent scenarios with graduated hints rather than giving you the answer directly. This mirrors the diagnostic reasoning the exam rewards.
• Bayesian Knowledge Tracing with a 0.90 mastery threshold, so the platform only advances you past a concept when your demonstrated accuracy justifies it.

For the agentic architecture domain specifically, the Agentic Architecture & Orchestration concept library covers every task statement in Domain 1, including the session management, decomposition, and hook patterns that appear most frequently in scenarios.

What is the fastest study path if you have two weeks?

A two-week plan that prioritises by exam weight:

The CCAR-F credential is valid for 12 months from the date it is awarded, so there is no benefit to delaying your attempt once you are consistently passing practice exams.