Tools and MCPs

Question 1 · 8.1

An agent has search_docs (“searches documents”) and search_kb (“searches the knowledge base”). It picks the wrong one roughly a third of the time. A stronger system prompt listing which to use when has not helped.

What is the most effective change?

• ASet tool_choice to force search_kb, which handles most queries correctly.
• BRewrite both descriptions to state their content, their boundary and what each is not for.
• CMerge them into one tool that searches both sources and returns combined results.
• DAdd examples of correct tool selection to the system prompt.

Question 2 · 8.1

During a payment provider outage, an inventory agent's check_stock tool raises a connection timeout. The exception propagates out of the tool loop and the user's session ends with a generic error page.

What should the harness do instead?

• ARetry the tool indefinitely until the provider recovers, so the agent eventually completes the task.
• BReturn a normal tool_result containing an empty stock list so the agent can continue.
• CRetry with backoff, then return a tool_result with is_error explaining that stock lookup is unavailable.
• DRemove check_stock from the tool list for the rest of the session so the model stops calling it.

Question 3 · 8.1

A team is designing an agent that must query an internal Postgres database inside their VPC, search the public web for supplier news, and email a summary to a manager.

Which two statements about tool implementation are correct? (Select 2.)

• AThe database query must be a client tool, because only their application can reach the VPC.
• BSending the email should be gated by approval or split into a request the agent cannot execute alone.
• CWeb search must also be a client tool, since it reaches systems outside Anthropic.
• DA single run_sql tool is preferable, since it covers any future query the agent may need.
• ESetting tool_choice to any will make the agent use each of the three tools in the right order.

Question 4 · 8.1

A finance agent can issue supplier payments. The team's proposal is a detailed system prompt stating that payments above 5,000 euros require sign-off, and an instruction to ask the user before calling payment_create.

What is the strongest objection to this design?

• AThe threshold should be enforced by the dispatcher or by not exposing an executing tool at all.
• BThe system prompt will be too long once every threshold rule is included.
• CThe agent may ask for confirmation too often, which will annoy finance staff.
• Dpayment_create should use strict so the amount always matches the schema.

Question 5 · 8.2

A team's MCP server works when tested with a script that pipes JSON to it, but the host reports malformed messages intermittently. The server uses a third-party HTTP library that prints connection-pool warnings, and it is launched as a subprocess.

What is the cause?

• AThe messages exceed the line-length limit of the stdio transport and are being split.
• BLibrary warnings are going to standard output, which must carry only valid MCP messages.
• CThe server is missing the initialize handshake, so the client rejects later messages.
• DThe subprocess is being restarted between requests, losing the session state.

Question 6 · 8.2

A documentation team wants Claude to be able to read their style guide when a writer asks for it, run a “draft release note” workflow that a writer starts by name, and look up a ticket's status whenever it becomes relevant mid-conversation.

How should these three be exposed?

• AAll three as tools, so the model can use whichever it needs.
• BStyle guide as a resource, release-note workflow as a prompt, ticket lookup as a tool.
• CStyle guide as a prompt, release-note workflow as a resource, ticket lookup as a tool.
• DAll three as resources, with the host deciding when to include each.

Question 7 · 8.2

An engineering team has a working stdio MCP server used inside Claude Code. Product now wants the same capabilities available from their customer-facing service, which calls the Messages API directly.

Which two statements are correct? (Select 2.)

• AThe server must be redeployed behind Streamable HTTP, since the connector accepts remote HTTPS servers only.
• BAny prompts or resources the server offers will not be reachable through the connector.
• CThe connector can launch the stdio server if the command is supplied alongside the URL.
• DMoving to HTTP requires re-implementing the tool handlers, since the semantics differ per transport.
• EThe protocol's session header will carry the customer's identity across calls.

Question 8 · 8.2

A developer is writing an MCP server that sometimes needs a missing value from the user — a cost centre it cannot infer — before it can complete a create_request call.

How should the server obtain it?

• ASend an elicitation request from the server to the client and wait for the reply.
• BFail the call with an error and rely on the model to ask the user and call again.
• CReturn an InputRequiredResult listing what it needs; the client retries with the responses.
• DOpen a subscriptions/listen stream and push a question to the client over it.

Question 9 · 8.3

A compliance team's Claude assistant produces review memos that vary in structure between analysts. Everything it needs to read is already available through tools it has. The team wants consistency and expects the format to evolve monthly.

What should they build?

• AAn MCP server exposing a generate_memo tool that renders the house format.
• BA skill holding the house format, definitions and an example memo, committed to the repository.
• CA longer system prompt containing the full format and all the definitions.
• DA set of custom tools, one per memo section, that the agent calls in order.

Question 10 · 8.3

A retailer's warehouse database is queried by three separate Claude-based applications. Each team has written its own query integration, and the schema is about to change.

Which two things make an MCP server the right move here? (Select 2.)

• ASeveral independent applications need the same capability.
• BA schema change can then be absorbed once, behind the server's interface.
• CMCP servers are faster than in-process tool handlers.
• DOnly an MCP server can require approval before a query runs.
• EThe Messages API cannot call custom tools, so a server is required.

Question 11 · 8.3

A developer wants their agent to summarise public vendor documentation pages and then run a short Python calculation on the numbers it finds. They are considering writing a scraper tool and a sandboxed execution service.

What is the best advice?

• AWrite the scraper but use a skill for the calculation, since the method must be consistent.
• BBuild both, so the behaviour is fully under their control and not subject to pricing changes.
• CUse the built-in web fetch and code execution tools, which already cover both needs.
• DUse an MCP server from the community that offers scraping and execution together.

Question 12 · 8.3

A team has a skill that documents their deployment procedure. It instructs the agent never to deploy on a Friday. Twice, the agent has deployed on a Friday anyway.

What is the correct conclusion?

• AA skill is advisory; the restriction belongs in a permission rule or the deploy tool itself.
• BThe skill's description should be rewritten so the agent loads it more reliably.
• CThe procedure should be moved into an MCP server so it is enforced by the protocol.
• DThe instruction should be repeated in the system prompt as well as the skill.