Claude Certified Architect - Foundations (CCAR-F; formerly CCA-F)

A team is building a customer-support agent. They want the simplest reliable architecture and only plan to add complexity if needed. According to Anthropic's guidance on building effective agents, what should they start with?

• AA multi-agent swarm where dozens of agents negotiate the answer
• BThe simplest solution possible, such as a single LLM call with retrieval and in-context e…
• CAn orchestrator-worker system with at least five specialized workers from day one
• DA fine-tuned model dedicated to support before writing any prompts

An architect needs to classify incoming requests and send each to a specialized prompt: billing questions to one prompt, technical issues to another, and general questions to a third. Which workflow pattern best fits this requirement?

• AParallelization
• BEvaluator-optimizer
• CPrompt chaining
• DRouting

A coding task requires generating a draft, then critiquing it against quality criteria, then refining it, repeating until the critique passes. Which pattern describes this, and what is the key design principle?

• AOrchestrator-worker, with one orchestrator spawning all critics at once
• BRouting, where a classifier picks the best draft from several
• CEvaluator-optimizer, separating generation and evaluation into distinct LLM calls with an…
• DParallelization, generating many drafts and averaging them

A research agent must handle tasks where the number and nature of subtasks cannot be predicted in advance: a central model must dynamically break the work down, delegate pieces, and synthesize the results. Which pattern is the best fit?

• APrompt chaining with a fixed sequence of steps
• BOrchestrator-worker
• CSectioning a single prompt into headers
• DRouting to one of three static categories

An architect wants to distinguish a 'workflow' from an 'agent' in Anthropic's terminology. Which statement is correct?

• AWorkflows let the LLM dynamically direct its own process and tool usage; agents follow pr…
• BWorkflows and agents are identical terms with no practical difference
• CAgents never use tools, while workflows always do
• DWorkflows orchestrate LLMs and tools through predefined code paths; agents dynamically di…

A latency-sensitive content pipeline must run a translation, a sentiment analysis, and a toxicity check on the same input, all independent of each other, then combine the results. Which pattern minimizes end-to-end latency?

• APrompt chaining the three checks sequentially
• BParallelization (sectioning), running the three independent calls concurrently and aggreg…
• CEvaluator-optimizer looping over each check
• DA single giant prompt asked to do all three plus self-grade

An agent autonomously calls tools in a loop. Anthropic recommends giving the agent guardrails and a way to stop. Which design element most directly bounds runaway loops in a production agent?

• ASetting a maximum number of turns/iterations and clear stopping conditions
• BIncreasing max_tokens to the model's maximum on every call
• CDisabling all tool use after the first call
• DRemoving the system prompt to reduce token cost

When designing a multi-step agent, Anthropic emphasizes 'showing the agent's planning steps' and keeping the design simple. What is the primary architectural benefit of this transparency?

• AIt guarantees the agent never makes a mistake
• BIt eliminates the need for any evaluation or testing
• CIt makes the agent's behavior easier to debug, audit, and trust, and easier to course-cor…
• DIt allows the model to skip tool calls entirely

A team uses the Claude Agent SDK and spawns several subagents in parallel for a large refactor. Each subagent does deep analysis. What is the main architectural advantage of using separate subagents here?

• ASubagents share one context window, so all analysis is visible to every agent
• BSubagents automatically inherit all parent permissions, removing any approval prompts
• CSubagents cannot use tools, which makes them faster
• DEach subagent gets its own context window, so one subagent's exploration does not pollute…

An architect must choose a model for a high-volume, latency-sensitive classification step inside a larger agent, where most decisions are simple. Which model selection strategy aligns with Anthropic's guidance?

• AAlways use the largest Opus model everywhere for maximum quality
• BUse a fine-tuned model because base models cannot classify
• CUse a faster, cheaper model such as Haiku for the simple high-volume step, reserving larg…
• DRoute every step to a random model to balance load

Anthropic warns that frameworks can obscure what is actually happening. What is their recommended approach when deciding whether to adopt an agent framework?

• AAlways adopt the most popular framework regardless of need
• BStart by using the LLM API directly since many patterns are a few lines of code, and if y…
• CNever use any framework under any circumstances
• DUse a framework only if it hides all prompts from you

A document-processing pipeline must (1) extract structured fields, (2) validate them against business rules, and (3) only then summarize. Step 2 acts as a programmatic gate. What is this gate called in prompt chaining, and why use it?

• AA 'guardrail' or programmatic check between steps that catches errors early before they p…
• BA 'router' that randomly selects which step runs next
• CAn 'orchestrator' that spawns workers for each field
• DA 'parallel branch' that runs all three steps at once

An agent occasionally takes irreversible actions (deleting records). Following agent safety guidance, what is the most appropriate architectural control before such an action executes?

• ALet the agent act freely and audit the logs afterward
• BRequire a human-in-the-loop approval or an explicit permission check for high-impact, irr…
• CIncrease the thinking budget so the model is more careful
• DSwitch to a smaller model to reduce the chance of action

A team measures whether their agent actually improved after adding an evaluator-optimizer loop. According to Anthropic, what should anchor this decision?

• ASubjective impressions from a single demo run
• BA systematic evaluation (evals) measuring task success on representative cases before and…
• CThe number of tool calls the agent makes
• DWhether the agent uses the newest model