• A. The full conversation history of every sub-agent, replayed verbatim on resume.

  A full replay is the highest-fidelity but least context-efficient option; it refills the window with material the agents do not need.

• B. Each sub-agent's final structured result to a known location; on resume the coordinator reloads those and re-dispatches only the unfinished work. (correct)

  Compact per-agent results keep fidelity while letting the coordinator re-dispatch only what is unfinished. This is the orchestrator plus compact-artifact pattern.

• C. Nothing; restart the whole pipeline from the first document.

  Restarting throws away completed work and wastes time and tokens.

• D. Only the coordinator's last message, letting each sub-agent rebuild its state from scratch.

  Sub-agents cannot reliably reconstruct prior findings from one message; you lose state.

Why: Persisting compact, structured results per agent preserves prior findings and lets the coordinator resume only the unfinished work, which balances fidelity against context cost.

• A. Add "return only JSON" to the prompt and parse whatever comes back.

  Prompt wording alone does not guarantee shape; it fails on the edge cases that matter.

• B. Define a strict output schema or tool the model must call, then validate the result and retry on a schema failure. (correct)

  A declared schema or tool call plus validation makes extraction contract-driven and self-correcting.

• C. Extract from the first brace to the last brace with a regular expression.

  Brace-matching breaks on nested or partial objects and silently returns wrong data.

• D. Set temperature to zero and accept the output as-is.

  Lower temperature reduces variation but does not enforce a structure or catch malformed output.

Why: Treating extraction as a validated contract (schema or tool plus a retry) is far more reliable than prompt-only or regex parsing.

• A. Estimate the refund status so the conversation keeps moving.

  Guessing risks telling the customer something false about their money.

• B. Tell the customer it is handled and reconcile it later.

  A false assurance breaks trust the moment it turns out to be wrong.

• C. Say plainly that the billing system is temporarily unavailable, give a realistic next step or timeframe, and offer to escalate. (correct)

  Honest failure handling with a concrete next step and an escalation path preserves trust.

• D. End the chat and tell the customer to try again later.

  Dropping the customer with no path forward is a poor outcome and avoidable.

Why: When a tool fails, stating the limitation honestly with a next step and an escalation route is better than fabricating a status or giving false assurance.

• A. Read every file in the repository into context.

  Bulk-reading floods the context window and buries the relevant lines.

• B. Search for the flag's identifier across the codebase, then open only the files that reference it. (correct)

  A targeted search narrows the surface before reading, keeping context focused on what matters.

• C. Open the README and guess the likely location.

  A guess from docs is unreliable and often points at stale information.

• D. Ask the user to paste the entire codebase.

  This shifts work to the user and still overflows context.

Why: Searching by the known identifier first, then reading only the matching files, is the context-efficient way to locate enforcement in a large codebase.