{
  "lesson": "16-multi-agent-orchestration-and-delegation",
  "title": "Multi-Agent Orchestration and Delegation",
  "questions": [
    {
      "stage": "pre",
      "question": "When is a subagent more appropriate than a simple tool?",
      "options": [
        "When a long reasoning task shares all context, permissions, and evaluation with the parent",
        "When the task needs its own bounded reasoning loop, evidence, context, and stop condition",
        "When a deterministic transformation needs an independent log and retry policy",
        "When an ambiguous task has no output contract or completion criteria yet"
      ],
      "correct": 1,
      "explanation": "A subagent earns a separate context for bounded reasoning, isolation, specialization, or independent review. Deterministic operations belong in code or tools."
    },
    {
      "stage": "check",
      "question": "Which requirement should orchestration code enforce rather than a coordinator prompt?",
      "options": [
        "Which ambiguous claim needs another evidence-gathering pass before synthesis",
        "Which authoritative source more directly supports a disputed semantic claim",
        "Cross-file review cannot start until required per-file results are complete or explicitly partial",
        "Whether one explanation better satisfies a qualitative clarity rubric"
      ],
      "correct": 2,
      "explanation": "Stage prerequisites are stable invariants. Code should track the manifest and block illegal transitions, while Claude handles semantic judgment within allowed states."
    },
    {
      "stage": "check",
      "question": "What should a partial subagent result contain?",
      "options": [
        "A complete-looking artifact with uncertainty notes but no explicit coverage boundary",
        "The full reasoning transcript and tool history without a normalized result contract",
        "A failure flag, the last attempted action, and a request for the coordinator to retry",
        "Valid completed work plus named gaps, failed evidence, and safe next actions"
      ],
      "correct": 3,
      "explanation": "Partial is a first-class state. It preserves usable work while preventing the coordinator from treating missing evidence as completion or negative evidence."
    },
    {
      "stage": "check",
      "question": "Why should an independent reviewer receive a fresh context?",
      "options": [
        "To reduce inheritance of the generator's framing and unsupported assumptions",
        "To let the reviewer judge intuitively without being constrained by the rubric",
        "To prevent the reviewer from seeing any evidence used by the generator",
        "To give the reviewer a larger independent token budget than the generator received"
      ],
      "correct": 0,
      "explanation": "Context isolation helps the reviewer judge artifact and evidence against the rubric instead of continuing the generator's persuasive narrative."
    },
    {
      "stage": "post",
      "question": "Three source researchers can work independently. What is the strongest orchestration?",
      "options": [
        "Run them sequentially so later agents inherit all prior framing and sources",
        "Run bounded tasks in parallel, normalize results, preserve failures, then reduce",
        "Give every researcher the same write-capable tools so any one can finish missing work",
        "Let each researcher draft the full report, then choose the most comprehensive version"
      ],
      "correct": 1,
      "explanation": "Independent work can reduce latency through parallel fan-out, but reliable reduction requires shared schemas, source identity, error propagation, and conflict rules."
    },
    {
      "stage": "post",
      "question": "A multi-agent design costs more without improving results. What should the architect do first?",
      "options": [
        "Keep the topology but evaluate only final answers so coordination costs do not dominate",
        "Add a coordinator agent to reduce duplicate work before changing the current roles",
        "Compare against the single-agent baseline and remove contexts without measured isolation, independence, or parallelism value",
        "Increase every agent's context and reasoning budget to recover quality lost to coordination"
      ],
      "correct": 2,
      "explanation": "Each context must justify its coordination cost with a measurable benefit. Removing unnecessary agents usually improves latency, cost, and debuggability."
    }
  ]
}