Agents and Workflows

Question 1 · 1.1

A retailer's support tool answers questions about deliveries. Every request is one of six intents, each with a documented resolution path, and volume is 20,000 a day. The team proposes a manager agent that spawns a subagent per request.

What is the strongest objection to the proposal?

• ASubagents cannot call external APIs, so every delivery lookup would fail.
• BKnown intents route to fixed paths more cheaply and predictably.
• CA manager agent cannot handle 20,000 requests a day at any price.
• DSubagents would each need their own API key, which is a security problem.

Question 2 · 1.1

A market-research team runs a lead agent that spawns subagents to investigate different aspects of a question in parallel. Costs are roughly fifteen times a plain chat, and leadership wants that justified.

Which two statements correctly describe what the extra spend buys? (Select 2.)

• AEach subagent explores in its own context window, so the system covers more ground than one sequential agent.
• BToken usage correlates strongly with performance on this kind of open-ended browsing work.
• CMulti-agent runs are deterministic, so results can be reproduced for audit.
• DSubagents return their full transcripts, giving the lead agent complete evidence.
• EParallel subagents reduce total token consumption compared with one agent doing the work.

Question 3 · 1.1

A platform team is refactoring a service that renames variables across a codebase. The change is mechanical, the files are known up front, and the current single agent occasionally edits the wrong file.

Which redesign best matches the guidance on architecture selection?

• AAdd an evaluator-optimizer loop so a second model reviews every edit.
• BSpawn one subagent per file so mistakes are isolated to a single context.
• CDo the rename deterministically and call the model only where judgement is needed.
• DKeep the agent and raise the turn limit so it can re-check its own work.

Question 4 · 1.2

An agent at a logistics firm occasionally runs rm -rf inside a shared working directory. The team's proposed fix is a stronger system prompt plus an example of what not to do.

What should they do instead?

• AAdd a PreToolUse hook on Bash that inspects the command and returns deny.
• BMove the instruction from the system prompt to the first user message for more weight.
• CAdd a PostToolUse hook that logs destructive commands for review.
• DLower the effort level so the agent takes fewer risky actions.

Question 5 · 1.2

A healthcare startup wants an agent that works through long document-processing jobs asynchronously. Its compliance team requires that patient files never leave infrastructure the company controls.

Which deployment approach fits the constraint?

• AManaged Agents with the default Anthropic-managed cloud sandbox.
• BThe Claude Code CLI run interactively by an operator each night.
• CThe Agent SDK in the company's own container, or Managed Agents with a self-hosted sandbox.
• DA hand-written Messages API loop, because only custom code can be compliant.

Question 6 · 1.2

A developer's Agent SDK job prints an empty summary to users roughly once a day. The logs show the run ended with subtype of error_max_turns.

What is the correct interpretation and response?

• AThe model refused the task; the prompt needs rewriting to avoid the refusal.
• BThe API returned an error; the request should be retried with backoff.
• CThe context window overflowed; compaction should be disabled.
• DThe loop hit its turn cap unfinished; resume the session or report an incomplete run.

Question 7 · 1.3

A research agent delegates to a subagent with the prompt “continue the analysis we discussed”. The subagent returns a generic summary that ignores the three files the parent had already read.

What explains the behaviour?

• AThe subagent ran out of turns before it could read the files.
• BSubagent tool permissions defaulted to read-only, blocking file access.
• CA subagent starts with no parent history, so “we discussed” refers to nothing it can see.
• DThe parent summarised the subagent's output before returning it.

Question 8 · 1.3

An agent processes long support transcripts. By turn 60 requests are slow and expensive, and the biggest contributors in the trace are repeated document-retrieval results from early in the run.

Which change addresses the cause most directly?

• AClear old tool results at a token trigger, keeping the latest few.
• BShorten the system prompt and remove examples from it.
• CLower max_tokens so every response the model writes is shorter.
• DSwitch to a model with a larger context window and change nothing else.

Question 9 · 1.3

A team wants an agent to answer a follow-up question a day later without re-reading the twelve files it analysed, and also wants to explore a second approach without losing the first thread.

Which combination does this correctly?

• AStore the final answer in a database and prepend it to a fresh prompt each time.
• BResume the saved session_id, and fork it when exploring the alternative.
• CRun both approaches as subagents from a new session each morning.
• DDisable compaction so nothing is ever summarised away.