Tool Design & MCP Integration

Question 1 · 2.1

A market-research agent has get_web_results (“Gets results for a query”) and get_article_data (“Gets data for a query”). The first searches the web; the second queries a licensed news archive. Logs show news questions going to web search about half the time.

What is the most effective first change?

• AAdd six routing examples to the system prompt showing which tool fits which question.
• BRewrite both descriptions to say what each source covers and when to use the other.
• CPut a lightweight classifier in front of the agent to pick the tool before Claude runs.
• DMerge both into one search tool that takes a free-text source parameter.

Question 2 · 2.1

A claims agent uses one tool, process_document, with a free-text task input. Depending on the wording, it returns extracted fields, a summary or a coverage opinion, and the downstream system that expects fields often breaks.

Which two changes address the design problem? (Select 2.)

• ASplit it into separate tools for extraction, summarising and coverage checks.
• BGive each new tool a description stating when to use it versus its siblings.
• CSet strict: true so the tool’s output always matches a schema.
• DTell Claude in the system prompt to always ask for extracted fields.
• EUse tool_choice to force process_document on every turn.
• FRaise max_tokens so every result has room for all three outputs.

Question 3 · 2.1

A retailer’s support agent has a well-described billing_issue_refund tool that says it is only for refunds already approved. The system prompt also says: “Whenever the customer mentions a refund, call billing_issue_refund.” Customers asking about the refund policy trigger refund attempts.

What should the architect do?

• AExpand the tool description further so it repeats the approval rule.
• BSet tool_choice to none whenever the message contains the word “policy”.
• CRemove the keyword rule and let intent-level guidance plus the description decide.
• DAdd a confirmation step before every refund, while keeping the rule in place.

Question 4 · 2.1

An internal IT agent connects to five MCP servers exposing about 60 tools. Several servers offer a tool called search or create_ticket, and Claude often files tickets in the wrong system.

Which change most directly fixes the selection problem?

• AList every tool and its server in a long table in the system prompt.
• BRemove the descriptions to save context, since the server name is enough.
• CMove to a larger model that can keep track of more tools at once.
• DGive tools service-prefixed names and descriptions naming the system they act on.

Question 5 · 2.2

An airline rebooking agent calls rebook_flight. When a fare rule forbids a change, the tool returns isError: true with the text “Operation failed”. The agent retries the same call four times, then tells the passenger the system is down.

What change best fixes this behaviour?

• ATell the agent in its system prompt never to retry rebook_flight calls at all.
• BFlag it as a non-retryable business error with a passenger-safe reason.
• CStop setting isError, so the agent treats the text as an ordinary result.
• DReturn a JSON-RPC protocol error instead of a normal tool result.

Question 6 · 2.2

A CRM lookup tool wraps every call in a try/except that returns an empty list on any exception. During a network outage, a sales agent tells several account managers that their key accounts have no open opportunities.

What is the core design flaw?

• AThe tool should return more fields per opportunity.
• BThe agent should double-check every empty result by calling again.
• CThe tool needs a longer timeout so the outage is less likely to matter.
• DAn access failure and a valid empty result look the same to the agent.

Question 7 · 2.2

In a research system, a document subagent hits a rate limit on one source and a permission error on another. It currently returns “Unable to complete task” to the coordinator and discards the eight documents it had already analysed.

How should the subagent handle these failures?

• ARetry the rate limit locally; escalate the permission error with the eight results.
• BKeep retrying both calls until they succeed, however long that takes.
• CSend both errors straight to the coordinator and let it do the retrying.
• DLeave out the failed sources and report success with the eight results.

Question 8 · 2.3

A research system’s synthesis agent has the same 14 tools as the search agent. Reports increasingly cite sources the search agent never found, and costs are rising.

What is the best change?

• ATell the synthesis agent in its prompt not to search unless necessary.
• BSet tool_choice to none for every synthesis request.
• CRemove search tools from the synthesis agent and give it one narrow fact-check tool.
• DMove all agents to a larger model so they follow scope better.

Question 9 · 2.3

A document pipeline defines four extraction tools, one per document type. With the default tool_choice, some documents come back as plain-text summaries and the pipeline fails.

Which tool_choice fixes this while keeping the right schema per document?

• A{"type": "any"}
• B{"type": "tool", "name": "extract_invoice"}
• C{"type": "auto"} with a stronger prompt
• D{"type": "none"}

Question 10 · 2.3

A compliance agent in the Agent SDK must never run shell commands. The team lists Read, Grep and Glob in allowedTools and leaves Bash out, but logs show Bash calls still being attempted.

Which two changes actually keep Bash away from the agent? (Select 2.)

• APass a tools list naming only Read, Grep and Glob.
• BAdd the bare name Bash to disallowedTools.
• CAdd Bash to allowedTools with a comment saying not to use it.
• DSet tool_choice to any so only approved tools can be chosen.
• ETell the agent in its system prompt that Bash is forbidden.

Question 11 · 2.3

An insurance claims workflow must always run extract_claim_fields on the first turn, then let Claude decide between fraud-check, coverage and payout tools based on what it finds.

How should tool_choice be configured?

• AForce extract_claim_fields on every request in the conversation.
• BUse any on every request so a tool is always called.
• CLeave auto and put “extract first” in the system prompt.
• DForce extract_claim_fields on the first request, then use auto.

Question 12 · 2.4

A team wants every developer on a repository to have the same Jira MCP server in Claude Code. Each developer has their own Jira API token.

What is the best configuration?

• AEach developer adds the server with --scope user and their token.
• BCommit .mcp.json with the server and one team token in its env.
• CCommit .mcp.json with the server and ${JIRA_TOKEN} in its env.
• DAdd the server at local scope and share the resulting ~/.claude.json.

Question 13 · 2.4

A developer is trying out a new MCP server for a code-metrics service. They are not sure the team will want it, and it needs their personal API key.

Where should they add it?

• ALocal scope, the default for claude mcp add.
• BProject scope, so the team can see it early.
• CProject scope with the key hard-coded, removed later.
• DA new ~/.claude/mcp.json file in their home folder.

Question 14 · 2.4

An agent answering questions about an analytics warehouse spends its first eight tool calls listing tables and describing columns before it writes any query. The schema changes only weekly.

Which change most directly reduces those exploratory calls?

• ARaise MAX_MCP_OUTPUT_TOKENS so each listing returns more at once.
• BTell the agent in its prompt to guess table names first.
• CAdd a second database server so the agent has more tools to try.
• DExpose the schema as an MCP resource the agent can start from.

Question 15 · 2.5

A developer asks Claude Code to find every place a deprecated function, legacyAuth, is called in a large monorepo so the calls can be replaced.

Which tool should Claude use first?

• AGlob with the pattern **/*legacyAuth*.
• BGrep for legacyAuth across the repository.
• CRead each file in the src directory in turn.
• DEdit with replace_all: true on the main entry file.

Question 16 · 2.5

Claude tries to change retries: 3 to retries: 5 for the payments client only, but Edit fails: the string appears four times in config.ts, once per client.

What should Claude do next?

• ARetry the same Edit; the error is often temporary.
• BSet replace_all: true so the Edit goes through.
• CInclude the payments block’s surrounding lines in old_string.
• DDelete config.ts and ask the user to recreate it.

Question 17 · 2.5

Claude Code joins an unfamiliar codebase and is asked how a password-reset email gets sent. There are around 3,000 files.

Which two steps reflect an effective approach? (Select 2.)

• AGrep for a distinctive string such as the reset email’s subject line.
• BRead the hit, then follow its imports and Grep for the functions it calls.
• CRead all files under src/ first so nothing is missed.
• DUse Glob with **/* to list every file before deciding.
• EWrite a new email module, since tracing the old one is slow.

Question 18 · 2.5

A team wants to find all test files for React components. Tests are named like Button.test.tsx and live next to their components throughout src/.

What is the most direct call?

• AGrep for describe( in all .tsx files.
• BRead src/index.tsx and follow its imports.
• CBash with ls src and inspect the output.
• DGlob with the pattern src/**/*.test.tsx.