An insurance claims agent is connected to three MCP servers exposing 95 tools in total. Logs show it often calls lookup_policy_v2 when get_policy was the right choice, and first-token latency has crept up. Only 14 tools appear in a month of successful runs.
A reporting agent must query sales data and occasionally look up a rarely used currency-conversion tool. Its MCP server also exposes delete_report and bulk_update_prices, which the agent never needs.
Which two configuration changes address the bloat correctly? (Select 2.)
In Claude Code, a team defines a log-summariser subagent with a name, a description and a prompt, but no tools field. The project also has a production database MCP server configured.
A coordinator delegates to eight specialist subagents. Two pairs have near-identical descriptions and delegation is inconsistent. What is the best first change?
A telecom support agent has a get_account tool that takes a customer_id argument and queries billing with a privileged service account. In testing, a message pasted into the chat convinces the agent to look up a different customer’s account.
A platform team runs the Agent SDK in CI to triage pull requests from external contributors. The container has a GitHub token and a cloud API key in environment variables so the agent can comment and query build logs.
An organisation’s managed Claude Code settings include the deny rule Bash(aws *). A project’s own settings add the allow rule Bash(aws s3 ls) so developers can list buckets.
A wealth-management firm’s advisers use a chat assistant that answers from a 30,000-token product guide. Answers are accurate, but advisers complain that nothing appears for three seconds. Total answer time is acceptable to them.
Which change best addresses the complaint without risking accuracy?
An architect must justify the configuration for a customer-facing claims-status agent to a review board. Three candidates have been tested: Opus 5 at high, Sonnet 5 at medium, and Haiku 4.5.
Which two pieces of evidence most strengthen the justification? (Select 2.)
A research team’s agent runs Opus 5 at max effort and takes 20 minutes per report. Reports are read the next morning, and the finance team wants lower cost without lower quality. What is the best change?
A payments company’s dispute assistant must return a structured decision within 8 seconds at p95. Sonnet 5 at high meets accuracy but has a p95 of 11 seconds, mostly thinking time. Sonnet 5 at low has a p95 of 5 seconds but fails multi-currency disputes, 6% of traffic. On the eval, medium effort handles those correctly with a p95 of 7 seconds.
A travel company’s booking assistant has 99.9% availability and stable p95 latency. Yet the share of conversations ending in a completed booking has fallen by a quarter since a prompt change two weeks ago.
Which monitoring addition would most directly have caught this earlier?
An insurer’s claims agent occasionally approves a claim it should have escalated. When engineers re-run the same claim, the agent escalates correctly every time.
A hospital group wants detailed monitoring of a clinical-documentation agent built on the Agent SDK. Compliance has not approved storing patient content in the observability platform. What is the best configuration?
A software company serves 2,000 business customers from one Claude deployment. The platform team proposes adding a customer ID and a session ID as labels on every metric so that finance can see cost per customer.
A retailer’s product assistant answers “Is the TrailMax 44-B7 tent waterproof?” with details of a different tent. The product page for 44-B7 is in the corpus. The index is embeddings-only, built from 1,000-token fixed-size chunks.
An insurer’s claims handbook is 120,000 tokens, changes once a quarter, and every question may touch any section. A team proposes a vector database, a chunker and a reranker.
A bank’s research assistant indexes analyst reports. Retrieval tests show chunks such as “Margins widened 40bps year on year” are found for the wrong company and period. The team wants the best improvement to recall without re-architecting.
Which two changes should the architect prioritise? (Select 2.)
A finance team’s assistant uses RAG over exported expense claims, one claim per chunk. When asked “What did Marketing spend on travel in Q2?” it gives a figure that is far too low, citing six claims.
An engineering organisation wants an assistant that answers questions about a 4-million-line monorepo that changes hundreds of times a day. A vendor proposes nightly embedding of every file.
Which approach best fits the data and query pattern?
A consultancy’s research assistant must answer questions such as “Summarise the AI regulation stance of each of the 27 EU member states’ data protection authorities.”
Which retrieval design fits this question pattern best?
A support bot answers 50,000 simple “how do I…” questions a day from a stable help centre. A team proposes making every question run through an agentic, multi-step search. What is the strongest objection?
A bank’s platform team wants one way for five internal AI assistants — built on Claude Code, an IDE extension and two custom agents — to search its internal knowledge base with the user’s own permissions.
A logistics company’s agent must book customs clearance through a broker. The broker runs its own agent that gathers documents, may ask follow-up questions about goods classification, and can take hours to finish. It will not expose its internal systems.
A team plans to call a vendor’s remote MCP server using the Claude API’s MCP connector. Which two constraints should they check before committing? (Select 2.)
A DevOps agent runs in a sandboxed container with a shell. It needs to list recent failed deployments and read their logs. A colleague proposes building a new MCP server wrapping the deployment platform’s API, which already ships a mature, well-documented CLI.
An internal IT agent has grown to 210 tools across eleven MCP servers. Tool definitions now use most of each request, and the agent often calls a similarly named tool from the wrong server. About six tools handle most tickets.
A healthcare provider’s assistant has a short rule set on handling patient identifiers that applies to every conversation. To reduce tokens, an engineer moves it into a skill with the description “PHI guidance”.
A contracts team uses Claude to check each 35-page master services agreement for internal inconsistencies, such as a liability cap in one clause contradicting an indemnity elsewhere. A proposal suggests indexing clauses and letting the agent look up only the clauses it thinks are relevant.
A team changes the contents of its tools array on every request, including only the tools a router predicts will be needed. Costs rose after the change. What is the most likely explanation?
You can change answers until you check. Nothing is saved or sent anywhere.