Browse all 10 free Claude Certified Architect - Foundations practice questions below.
Your support agent's loop calls Claude with the support tools and tool_choice "auto". If the response contains a text block, the loop sends that text to the customer and ends the turn. Otherwise it runs each requested tool, appends the results and calls Claude again. A refund usually takes three or four tool calls.
Production logs show that 7% of conversations stall after the agent says "Let me pull up that order for you." Each of those responses had two content blocks: that sentence, then a lookup_order tool_use block.
What change most effectively fixes the stalls?
- Add a system prompt rule that tells Claude to call tools without commentary and to save customer-facing text for its final answer.
- Add a classifier call that labels each text block as a final answer or a progress update, and keep looping on progress updates.
- Set tool_choice to "any" on the first request after each customer message, so Claude calls a tool before it writes any text.
- Replace the text check with stop_reason: on "tool_use", run the tools and return results; on "end_turn", reply to the customer.
Your multi-agent research system uses a hub-and-spoke architecture where a coordinator agent dispatches tasks to specialized subagents (market research, patent search, competitor analysis). A subagent needs context about the overall research objective.
What is the correct way to provide this context?
- Persist the research objective in a shared key-value store that every subagent reads from on cold start
- Pass complete findings and the research objective explicitly in the prompt sent to each subagent
- Rely on automatic context inheritance, since subagents receive the coordinator's message history by default
- Have each subagent invoke the coordinator via a callback tool to fetch any context it requires mid-run
Your support agent issues refunds through a process_refund tool. Company policy says refunds above $500 need a manager's approval. In a red-team test, 3 of 200 conversations ended with a refund between $520 and $900, each after the customer claimed a supervisor had already approved it.
Which change makes the $500 limit hold in every conversation?
- Check the amount inside the refund tool and, above $500, return an error telling the agent to escalate to a manager.
- Set "maximum": 500 on the amount field of the refund tool's input schema, so the model cannot request a larger refund.
- Have a second Claude call review each refund decision against the policy text before the refund tool runs.
- State the $500 limit in the system prompt and add few-shot examples that show the agent escalating large refunds.
Your research coordinator delegates patent search, market analysis and competitor review to three specialized subagents. The tasks are independent, and each takes about four minutes. Traces show the coordinator sends the patent task, waits for its result, then sends the next one, so every report takes about twelve minutes.
Which change brings a report down to about four minutes while keeping the coordinator in charge of synthesis?
- Start three coordinator sessions with one subagent each, then merge their reports in application code afterward.
- Submit the three subagent prompts through the Message Batches API so they are processed together at half the cost.
- Have the coordinator request all three subagents as Task tool calls within one response, instead of across turns.
- Give one subagent all three research briefs so that it can work through them together in a single context window.
Your customer support agent uses Claude's self-reported confidence scores to decide when to escalate to a human agent. During a production review, you discover that the agent occasionally processes complex policy-edge-case tickets without escalating, despite producing incorrect resolutions.
What is the MOST likely root cause?
- The prompt lacks explicit, numbered instructions describing exactly when the agent should report low confidence
- Self-reported confidence scores are poorly calibrated and should not be used as the primary escalation mechanism
- The model was never fine-tuned on edge-case escalations, so its confidence scores remain uncalibrated for those tickets
- The escalation threshold is set too low, so tickets need to report very low confidence before they ever reach a human
Your billing agent's loop continues while stop_reason is "tool_use" and treats every other stop_reason as a finished turn, sending the text to the customer. Most conversations work, but about 4% of answers to multi-item refund questions end mid-sentence, for example "Your refund for the first two items will be".
What change fixes this?
- Resend the request whenever the reply does not end with a period, until a complete answer comes back from Claude.
- Add a system prompt rule asking Claude to keep each reply under 350 tokens, so that answers fit inside the limit.
- End the turn only on "end_turn", and when stop_reason is "max_tokens", retry with a higher limit before replying.
- Trim any incomplete final sentence from the reply before sending it, so that customers see whole sentences.
Your coordinator spends its first ten turns clarifying a research request with the user, who says the analysis must cover only the European market and exclude two competitors that are being acquired. The coordinator then dispatches a market-analysis subagent with the brief "Analyze the competitive landscape for our product." The subagent's report covers North America and includes both excluded competitors.
What is the most likely cause?
- The coordinator's ten-turn conversation overflowed the subagent's context window, so the earliest constraints were dropped.
- The subagent received just the text of its Task prompt, and the coordinator never wrote the user's limits into it.
- The subagent's own system prompt overrode the inherited user constraints with its default instructions for broad analysis.
- The subagent inherited the conversation but weighed the brief more heavily than the user's earlier constraints.
Your support agent, built on the Claude Agent SDK, issues refunds through a process_refund tool owned by another team, and you cannot change that tool. Policy says refunds above $500 must go to a human reviewer. The system prompt states this in capital letters, yet audits find a few refunds between $600 and $800 every month.
Where should you enforce the limit?
- In the system prompt, by moving the rule to the end and adding three examples of the agent escalating large refunds.
- In the process_refund tool description, by stating that amounts above $500 are refused and must be escalated instead.
- In a PostToolUse hook that checks each process_refund result and flags any refund above $500 for a human reviewer.
- In a PreToolUse hook that denies process_refund calls above $500 and returns a reason telling the agent to escalate.
Your team shares two MCP servers through the repository's .mcp.json: a GitHub server that needs each developer's personal token, and a database tool whose launch command needs a socket path. Most developers use /var/run/pg.sock, but a few have it elsewhere. You want one committed file that works for everyone without anyone editing it locally.
What should .mcp.json contain?
- Both values written out in .mcp.json, with the file added to .gitignore so that each developer keeps a private copy.
- The token as ${GITHUB_TOKEN}, and the socket path written out as /var/run/pg.sock, with the few exceptions editing their copy.
- The token and the socket path as ${GITHUB_TOKEN} and ${PG_SOCKET}, with a README that asks everyone to set both variables.
- The token as ${GITHUB_TOKEN}, and the socket path as ${PG_SOCKET:-/var/run/pg.sock}, so only the exceptions set a variable.
Your monorepo has packages/billing (Python), packages/web (React) and packages/shared-ui (React). The platform team maintains company-wide standards files: python-style.md, react-style.md and security.md. Each package's maintainers want Claude Code to apply the standards relevant to their package, plus a few notes of their own, without loading unrelated standards.
What setup fits best?
- A root CLAUDE.md that @imports all three standards files, with headings that say which packages each file applies to.
- One .claude/rules/ file per standards file, with no paths frontmatter, so Claude picks the relevant rules for each edit.
- A CLAUDE.md in each package directory that @imports the relevant standards files and adds the package's own extra notes.
- One skill per standards file, with descriptions naming the packages, so Claude loads each one when it judges it relevant.