Claude Certified Developer - Foundations (CCDV-F)

A team is building a customer-support feature with Claude. They want the simplest reliable design and will only add agent complexity if needed. What should they start with?

• AA multi-agent swarm in which several agents debate every reply
• BA single well-prompted call with retrieval, adding steps if needed
• CAn orchestrator-worker system with five specialized workers from day one
• DA fine-tuned support model trained before any prompt is written

You need to classify each support ticket and send billing questions to one prompt, technical issues to another, and general questions to a third. Which workflow pattern best fits?

• AParallelization, drafting all three category answers concurrently
• BAn evaluator-optimizer loop that critiques each draft answer
• CPrompt chaining through billing, technical, and general steps
• DRouting, classifying the ticket to one specialized prompt

A coding assistant generates a draft, then a separate call critiques it against quality criteria, then refines it until the critique passes. Which pattern is this?

• AOrchestrator-worker, where a manager delegates subtasks to workers
• BRouting, dispatching the draft to a category-specific prompt
• CEvaluator-optimizer, looping generate and critique until it passes
• DParallelization, drafting independent sections concurrently

A research task can be split into independent section drafts that are later merged. Which pattern best fits, and what is a key risk?

• APrompt chaining, whose main risk is an unnecessarily long sequential token bill
• BParallelization by sectioning, risking inconsistent facts without a merge
• CRouting, whose main risk is misclassifying the whole document
• DEvaluator-optimizer, whose main risk is an unbounded critique loop

When is an orchestrator-worker agent pattern most appropriate?

• AWhen every request follows the same fixed three-step pipeline
• BWhen subtasks emerge at runtime and a manager must delegate
• CWhen the job only needs a single classification label returned
• DWhen the whole job already fits in one call without any tools

In the Claude Agent SDK, what best describes a typical agent loop?

• AOne non-streaming Messages call that never returns any tool_use blocks
• BA cycle of reasoning, tool calls, tool results, and continuation
• CA set of offline batch jobs with no interactive tool feedback
• DA fixed stored procedure whose steps the model cannot influence

A pipeline extracts entities with a fast model, enriches each entity with tools on a mid-tier model, then writes the final report on the strongest model. What is the main benefit?

• AIt guarantees fewer total tokens than one call to the strongest model
• BIt puts fast models on routine steps and the strongest on synthesis
• CIt removes the need to handle tool errors at the enrichment step
• DIt guarantees the final report will never need human review

A developer implements an agent that can call many tools. The agent sometimes loops indefinitely calling the same failing tool. What is the best first mitigation?

• ARemove every tool so that the model can only produce plain chat text
• BAdd max turns, repeated-failure detection, and clear tool errors
• CSend every request to the largest available model instead
• DTurn off streaming so partial output cannot be rendered

When should you prefer a fixed prompt-chaining pipeline over a free-form agent loop?

• AWhen the steps are known and stable and need no runtime planning
• BWhen the application happens to be running on the smallest model
• CWhen the available tool schemas change unpredictably per request
• DWhen the model is expected to define new tools during the run

An evaluator-optimizer loop is not improving quality after several iterations. What is the most likely design issue?

• AThe responses are being streamed rather than returned in one piece
• BThe evaluator lacks testable criteria or shares the generator's prompt
• CThe critique step was not submitted through the asynchronous Batches API
• DThe project CLAUDE.md file was not attached to the API request

You need both low latency for classification and deep reasoning for rare complex tickets. Which combined pattern fits best?

• AMajority voting across five identical calls to the strongest available model
• BRouting simple tickets to a fast model and escalating complex ones
• CRunning the full multi-agent stack on every incoming ticket
• DDisabling tool access specifically on the most complex tickets

In multi-agent designs, what is a primary reason to use specialized subagents instead of one mega-prompt that does everything?

• ASubagents reduce the total tokens a task consumes to nearly zero
• BNarrower instructions, tools, and success criteria per role
• CAnthropic's guidance prohibits shipping single-agent designs
• DSubagents make per-agent system prompts unnecessary to write

A content-moderation check is run three times on the same input with different prompts, and the verdicts are combined into one decision. Which parallelization variation is this?

• ASectioning, which splits one task into independent subtasks run at once
• BVoting, which runs the same task several times to get diverse outputs
• CRouting, which classifies the input and sends it down a single path
• DChaining, which feeds each step's output into the step that follows

What is a practical stop condition for a tool-using agent in production?

• AKeep looping until the end user closes the browser tab
• BA max-turn cap, a timeout, and a final message with no tool_use
• CStop only after the agent has issued exactly one hundred separate tool calls
• DStop as soon as the configured model identifier changes

You are deciding between a single tool-using agent and an orchestrator with three workers. The task is "summarize this PDF." What should you choose and why?

• AUse three workers, because multi-agent is the recommended default design
• BStart with a single call or agent; summarization rarely needs workers
• CUse workers only after switching off the Messages API for that route
• DUse Batch API workers so the chat interface stays responsive