{
  "schema": "insight.wwdc.priority-session.refined-zh.v1",
  "session_id": "299",
  "title": "Create robust evaluations for agentic apps",
  "title_zh": "Agent / 评估：Create robust evaluations for agentic apps",
  "priority": "P1",
  "themes": [
    "xcode_agents",
    "evaluations_observability",
    "local_ai_mlx_core_ai",
    "spatial_media_games"
  ],
  "theme_labels_zh": [
    "Xcode Agents 与开发工具",
    "评估与可观测",
    "端侧 AI / MLX / Core AI",
    "空间计算、媒体与游戏"
  ],
  "positioning_zh": "Agentic Apps 评估：把多步任务、工具调用和状态变化纳入鲁棒性测试。",
  "executive_takeaway_zh": "299 相比 298 更聚焦 agentic 场景，重点是多步任务的鲁棒性、可恢复性和边界条件。",
  "refined_claims": [
    {
      "id": "r1",
      "title_zh": "多步评估",
      "claim_zh": "Agentic app 不能只评估单轮输出，还要评估工具选择、步骤顺序和中间状态。",
      "evidence_refs": [
        "a1",
        "a2",
        "a3"
      ],
      "evidence_policy": "safe anchor ids only; no full transcript text",
      "review_status": "editorial_refined_v1_needs_quote_check"
    },
    {
      "id": "r2",
      "title_zh": "鲁棒性",
      "claim_zh": "鲁棒评估需要覆盖异常输入、工具失败、权限缺失和用户中断。",
      "evidence_refs": [
        "a2",
        "a3",
        "a4"
      ],
      "evidence_policy": "safe anchor ids only; no full transcript text",
      "review_status": "editorial_refined_v1_needs_quote_check"
    },
    {
      "id": "r3",
      "title_zh": "产品发布",
      "claim_zh": "如果 agent 不能稳定完成关键任务，就应限制能力范围或增加人工确认。",
      "evidence_refs": [
        "a3",
        "a4",
        "a5"
      ],
      "evidence_policy": "safe anchor ids only; no full transcript text",
      "review_status": "editorial_refined_v1_needs_quote_check"
    },
    {
      "id": "r4",
      "title_zh": "治理价值",
      "claim_zh": "评估结果应反过来定义权限、提示、工具 schema 和 fallback 策略。",
      "evidence_refs": [
        "a4",
        "a5",
        "a6"
      ],
      "evidence_policy": "safe anchor ids only; no full transcript text",
      "review_status": "editorial_refined_v1_needs_quote_check"
    }
  ],
  "evidence_path": {
    "brief": "data/session_briefs_zh/299.json",
    "deep_skeleton": "data/priority_session_deep_zh/299.json",
    "workbench": "workbench.html?view=sessions/299/session_299_workbench.html",
    "safe_anchors": [
      {
        "id": "a1",
        "time": 8,
        "label": "0:08",
        "sourceSegment": 0,
        "public_text_zh": "内部字幕片段 #0 · 仅公开时间锚点和安全摘要，不公开原文或完整中文翻译。",
        "apple_url": ""
      },
      {
        "id": "a2",
        "time": 266,
        "label": "4:26",
        "sourceSegment": 79,
        "public_text_zh": "内部字幕片段 #79 · 仅公开时间锚点和安全摘要，不公开原文或完整中文翻译。",
        "apple_url": ""
      },
      {
        "id": "a3",
        "time": 527,
        "label": "8:47",
        "sourceSegment": 158,
        "public_text_zh": "内部字幕片段 #158 · 仅公开时间锚点和安全摘要，不公开原文或完整中文翻译。",
        "apple_url": ""
      },
      {
        "id": "a4",
        "time": 780,
        "label": "13:00",
        "sourceSegment": 238,
        "public_text_zh": "内部字幕片段 #238 · 仅公开时间锚点和安全摘要，不公开原文或完整中文翻译。",
        "apple_url": ""
      },
      {
        "id": "a5",
        "time": 1016,
        "label": "16:56",
        "sourceSegment": 317,
        "public_text_zh": "内部字幕片段 #317 · 仅公开时间锚点和安全摘要，不公开原文或完整中文翻译。",
        "apple_url": ""
      },
      {
        "id": "a6",
        "time": 1277,
        "label": "21:17",
        "sourceSegment": 396,
        "public_text_zh": "内部字幕片段 #396 · 仅公开时间锚点和安全摘要，不公开原文或完整中文翻译。",
        "apple_url": ""
      }
    ],
    "policy": "safe_editorial_summary_only_no_full_transcript_or_full_translation"
  },
  "review_questions_zh": [
    "任务完成率如何定义？",
    "工具调用错误如何分类？",
    "失败时是否有安全退出路径？"
  ],
  "publication_guidance_zh": [
    "可用于汇报页、Session 介绍和主题报告入口。",
    "发布前仍需人工核对 Apple 官方表达和少量合规短引文。",
    "不得把完整 transcript 或完整中文翻译复制到公开页。"
  ],
  "quality": {
    "level": "P0/P1 editorial refined v1",
    "requires_human_quote_review": true,
    "requires_official_term_check": true,
    "generated_at": "2026-06-22"
  }
}