{ "generated_at": "2026-09-11", "description": "Evidence-linked question signals derived from curated public interview records. Frequency means distinct source records in this repository, not total interviews across the market.", "schema_version": 1, "items": [ { "id": "agent-evaluation-system", "question": "如何设计 Agent 的离线评测、线上指标和失败样本回流？", "category": "Agent Evaluation", "companies": ["字节跳动", "阿里巴巴", "美团", "米哈游", "蚂蚁国际"], "rounds": ["一面", "二面", "项目深挖"], "source_ids": ["nowcoder-bytedance-agent-roundup-2026-08-28", "nowcoder-alibaba-agent-roundup-2026-08-31", "xhs-meituan-agent-2026-06-18", "nowcoder-mihoyo-agent-first-round-2026-08", "nowcoder-ant-international-2025-09-24"], "source_count": 5, "last_seen_at": "2026-08-31", "source_types": ["first_hand", "aggregation"], "confidence": "high", "answer_outline": ["定义任务成功条件", "区分 outcome/trajectory/state grading", "构建失败与对抗样本", "建立回归门禁", "监控成本延迟和安全"], "evaluation_points": ["不能只用 LLM-as-Judge", "要区分模型自述完成与独立验证", "能说明数据污染和评测漂移"] }, { "id": "context-memory-compression", "question": "长上下文、Session、Working State 与长期 Memory 如何分层？", "category": "Context Engineering", "companies": ["字节跳动", "腾讯", "百度", "阿里巴巴"], "rounds": ["一面", "二面"], "source_ids": ["nowcoder-bytedance-agent-first-round-2026-09-03", "nowcoder-bytedance-agent-intern-2026-03", "xhs-tencent-ai-app-2026-06-29", "xhs-baidu-llm-app-2026-03-28", "nowcoder-alibaba-taotian-agent-intern-2026-04"], "source_count": 5, "last_seen_at": "2026-09-04", "source_types": ["first_hand"], "confidence": "high", "answer_outline": ["分离短期上下文与长期记忆", "按预算触发压缩", "保留证据和 Artifact 引用", "评估压缩后的任务成功率"], "evaluation_points": ["说明压缩失真风险", "说明恢复时如何重建上下文", "不能把所有历史都塞回 Prompt"] }, { "id": "tool-mcp-skill-boundary", "question": "Tool、Function Calling、MCP 与 Skill 的边界是什么？", "category": "Tooling", "companies": ["字节跳动", "阿里巴巴", "快手", "米哈游", "蚂蚁"], "rounds": ["一面", "二面"], "source_ids": ["nowcoder-bytedance-agent-intern-2026-03", "nowcoder-alibaba-taotian-agent-intern-2026-04", "nowcoder-kuaishou-agent-fullstack-second-round-2026-09", "nowcoder-mihoyo-agent-first-round-2026-08", "nowcoder-ant-agent-2026-05-20"], "source_count": 5, "last_seen_at": "2026-09-07", "source_types": ["first_hand"], "confidence": "high", "answer_outline": ["Function Calling 是模型输出约定", "Tool 是可执行能力", "MCP 是工具和资源接入协议", "Skill 是可复用工作方法与资源包"], "evaluation_points": ["能说明权限边界", "能说明渐进式加载", "能举出不该使用 MCP 的场景"] }, { "id": "harness-runtime-orchestrator", "question": "Agent、Orchestrator、Harness 和 Runtime 分别负责什么？", "category": "Agent Runtime", "companies": ["字节跳动", "小红书", "阿里巴巴"], "rounds": ["一面", "系统设计"], "source_ids": ["nowcoder-bytedance-agent-roundup-2026-08-28", "nowcoder-bytedance-agent-first-round-2026-09-03", "nowcoder-xiaohongshu-agentic-fullstack-2026-08", "nowcoder-alibaba-agent-roundup-2026-08-31"], "source_count": 4, "last_seen_at": "2026-09-04", "source_types": ["first_hand", "aggregation"], "confidence": "high", "answer_outline": ["Agent 负责语义决策", "Orchestrator 负责依赖和分派", "Harness 约束模型工具循环", "Runtime 管理生命周期和恢复"], "evaluation_points": ["边界清晰", "能解释为什么 while loop 不够", "能画出控制面与执行面"] }, { "id": "rag-failure-diagnosis", "question": "RAG 召回或回答效果差时，如何逐层定位？", "category": "RAG", "companies": ["字节跳动", "美团", "快手", "百度", "京东"], "rounds": ["一面", "项目深挖"], "source_ids": ["csdn-byte-agent-memory-rag", "xhs-meituan-agent-2026-06-18", "xhs-kuaishou-agent-2026-03-15", "xhs-baidu-llm-app-2026-03-28", "nowcoder-jd-agent-intern-2026-04-23"], "source_count": 5, "last_seen_at": "2026-06-18", "source_types": ["first_hand", "aggregation"], "confidence": "high", "answer_outline": ["解析和切块", "Query Rewrite", "Sparse/Dense 召回", "RRF/Rerank", "生成忠实度", "线上反馈"], "evaluation_points": ["有分层指标", "能定位表格和扫描 PDF", "区分召回失败与生成失败"] }, { "id": "multi-agent-necessity", "question": "为什么一定要使用 Multi-Agent，单 Agent 或 Workflow 不行吗？", "category": "Multi-Agent", "companies": ["字节跳动", "腾讯", "阿里巴巴", "快手"], "rounds": ["一面", "二面", "项目深挖"], "source_ids": ["xhs-bytedance-agent-second-round-2026-05-30", "xhs-tencent-ai-app-2026-06-29", "nowcoder-ali-ant-byte-agent-summary-2026-04-24", "xhs-kuaishou-agent-intern-2026-06-12"], "source_count": 4, "last_seen_at": "2026-06-29", "source_types": ["first_hand", "aggregation"], "confidence": "high", "answer_outline": ["先给单 Agent 基线", "证明上下文隔离或并行收益", "说明通信和一致性成本", "提供消融实验"], "evaluation_points": ["不能用角色数量代替架构理由", "说明失败传播", "给出不用 Multi-Agent 的条件"] }, { "id": "tool-retry-idempotency", "question": "工具调用失败或超时时，如何安全重试并避免重复副作用？", "category": "Reliability", "companies": ["腾讯", "阿里巴巴", "字节跳动"], "rounds": ["一面", "系统设计"], "source_ids": ["nowcoder-tencent-baidu-agent-summary-2026-04-28", "nowcoder-alibaba-agent-roundup-2026-08-31", "nowcoder-bytedance-agent-first-round-2026-09-03"], "source_count": 3, "last_seen_at": "2026-09-04", "source_types": ["first_hand", "aggregation"], "confidence": "high", "answer_outline": ["错误分类", "幂等键", "结果日志", "指数退避", "补偿事务", "人工审批"], "evaluation_points": ["认识到超时不等于未执行", "高风险操作不能自动无限重试", "恢复要读取执行日志"] }, { "id": "agent-path-oscillation", "question": "如何检测并处理 Agent 路径震荡和无进展循环？", "category": "Reliability", "companies": ["字节跳动"], "rounds": ["一面", "系统设计"], "source_ids": ["nowcoder-bytedance-agent-roundup-2026-08-28"], "source_count": 1, "last_seen_at": "2026-08-28", "source_types": ["aggregation"], "confidence": "medium", "answer_outline": ["动作与错误指纹去重", "进展检测", "失败预算", "强制重规划", "转人工"], "evaluation_points": ["区分探索和震荡", "不能只增加 max_iterations", "能定义 no-progress 信号"] }, { "id": "ai-coding-quality", "question": "如何验证 AI Coding 生成的代码，而不是直接接受结果？", "category": "Coding Agent", "companies": ["字节跳动", "快手", "阿里巴巴", "腾讯", "米哈游"], "rounds": ["一面", "二面", "现场编码"], "source_ids": ["nowcoder-bytedance-agent-first-round-2026-08-12", "nowcoder-kuaishou-agent-first-round-2026-09-03", "nowcoder-alibaba-taobao-agent-2026-04-30", "nowcoder-tencent-agent-app-2026-04", "nowcoder-mihoyo-agent-first-round-2026-08"], "source_count": 5, "last_seen_at": "2026-09-03", "source_types": ["first_hand"], "confidence": "high", "answer_outline": ["静态检查", "单元和属性测试", "安全扫描", "差异审查", "运行时验证", "回滚"], "evaluation_points": ["区分生成和交付", "能解释测试盲区", "保留证据而非口头确认"] }, { "id": "agent-cost-latency", "question": "包含多个模型和工具调用的 Agent 链路如何优化延迟与成本？", "category": "Performance", "companies": ["字节跳动", "阿里巴巴", "快手"], "rounds": ["一面", "项目深挖"], "source_ids": ["nowcoder-bytedance-agent-first-round-2026-09-03", "nowcoder-alibaba-taobao-agent-2026-04-30", "xhs-kuaishou-agent-2026-03-15"], "source_count": 3, "last_seen_at": "2026-09-04", "source_types": ["first_hand"], "confidence": "high", "answer_outline": ["Trace 分解延迟", "并行无依赖调用", "缓存", "模型路由", "上下文预算", "超时与降级"], "evaluation_points": ["给出 P50/P95", "不能只说换小模型", "同时报告质量回归"] }, { "id": "agent-training-reward", "question": "Agent 多步工具轨迹如何设计 Reward 和信用分配？", "category": "Agentic RL", "companies": ["阿里巴巴", "腾讯", "字节跳动"], "rounds": ["一面", "算法面"], "source_ids": ["nowcoder-taotian-agent-social-first-round-2026-07", "nowcoder-tencent-agent-app-2026-04", "nowcoder-bytedance-ant-tencent-agent-intern-2026-07"], "source_count": 3, "last_seen_at": "2026-07-22", "source_types": ["first_hand"], "confidence": "high", "answer_outline": ["Outcome Reward", "Process Reward", "格式与工具合法性", "延迟成本惩罚", "Reward Hacking 防护"], "evaluation_points": ["说明稀疏奖励", "说明 token/turn/trajectory 粒度", "区分训练指标和线上成功"] }, { "id": "structured-output-reliability", "question": "模型生成工具参数或结构化评分时，如何保证输出合法且语义一致？", "category": "Structured Output", "companies": ["字节跳动", "阿里巴巴", "美团"], "rounds": ["一面"], "source_ids": ["nowcoder-bytedance-agent-first-round-2026-09-03", "nowcoder-alibaba-taobao-agent-2026-04-30", "xhs-meituan-agent-2026-06-18"], "source_count": 3, "last_seen_at": "2026-09-04", "source_types": ["first_hand"], "confidence": "high", "answer_outline": ["JSON Schema", "约束解码", "语义校验", "修复重试", "失败降级"], "evaluation_points": ["格式正确不等于业务正确", "参数需要权限和范围校验", "保留原始输出用于审计"] } ] }