面试题库

30 天从前端工程师到 Agent 工程师

D16 Supervisor 动态路由：structured output 路由、override、routingReason

• 为什么要让模型输出 structured output 而不是自然语言来做路由？自然语言到底差在哪？Why use structured output rather than natural language for routing? What exactly goes wrong with free text?

  国内高频海外高频进阶#structured-output#routing#reliability

  分析过程 · 先想清楚再作答

  1. 这题最容易答成「结构化更规范、更好解析」——这是形容词，不是理由。面试官想听的是一个具体的失败场景，最好是你真的调过的那种。
  2. 把最锋利的一刀先亮出来：**自然语言路由的失败是静默的**。模型回「我觉得这个可以让查订单的同事看一下」，它其实判对了，但说的是人话不是 id，正则匹配不上就落进兜底，日志里只留下一个 smalltalk。模型是对的、解析是错的，而它和「模型判错了」在日志里长得一模一样。你会去调提示词，调两天才发现问题在那三行正则。
  3. 然后给三个漏洞，一条一条对上结构化输出解决了什么：输出会漂移（今天回「订单查询」明天回「查订单」，正则永远追不上，模型小版本升级你就掉准确率）；没有置信度（自然语言里没有「我有多大把握」这个信息，你没法区分它很确定还是在猜）；拼错或自造的路由名要到运行时才炸（枚举是一道编译期就存在的闸门）。
  4. 接着说清机制，别停在「用 zod 更规范」：把 schema 发进请求（response_format 里的 json_schema），模型的解码过程被枚举约束；回来之后**用同一份声明再校验一遍**。一份声明两用，请求与校验不会漂移。
  5. 关键的反直觉点，答到这里就拉开差距了：**结构化输出不等于不用校验**。不是所有网关、所有模型都严格执行 schema，降级到备用模型时更说不准。所以解析函数的返回类型应该是「一段待校验的东西」，而不是「已经是 RouteDecision」。
  6. 可以预期的追问：那不支持 json_schema 的模型怎么办？答案是退回「few-shot 加严格提示词加自己校验」，闸门仍然在你的枚举校验那一步——真正兜底的从来不是模型的自觉，是你的解析层。

  How to reason about it · think before answering

  1. The trap is answering structured output is cleaner and easier to parse. Those are adjectives, not reasons. The interviewer wants a concrete failure you have actually debugged.
  2. Lead with the sharpest point: free-text routing fails silently. The model replies I think the order desk should look at this — it judged correctly, but it spoke prose, not an id. Your regex misses, you fall through to the default, and the log shows only smalltalk. A correct model with a broken parser looks exactly like a wrong model, so you spend two days tuning a prompt that was never the problem.
  3. Then list three holes and map each to what structured output fixes: wording drifts across versions so regexes never catch up; there is no confidence signal, so you cannot tell certainty from guessing; and an invented route name only explodes at runtime, whereas an enum is a gate that exists before the request is even sent.
  4. Explain the mechanism rather than stopping at zod is nicer: send the schema in the request (response_format with a json_schema), so decoding is constrained by the enum, then validate the response with the same declaration. One declaration used twice means request and validation cannot drift apart.
  5. The counterintuitive point that separates candidates: structured output does not remove the need to validate. Not every gateway or model enforces the schema strictly, and a fallback model may not at all. Your parse function should return something-to-be-validated, not an already-typed decision.
  6. Expect: what if the model does not support json_schema? Fall back to few-shot plus a strict prompt plus your own validation. The real gate was never the model's discipline; it is your parsing layer.

  答题要点

  • 自然语言路由的失败是静默的：模型判对了但说的是人话，正则匹配不上就落兜底，和判错在日志里完全一样
  • 三个漏洞：措辞会漂移（正则追不上）、没有置信度（分不清确定与猜）、自造的路由名要到运行时才炸
  • 机制是一份声明两用：schema 随请求发出去约束解码，回来后用同一份声明校验，请求与校验不会漂移
  • 枚举是编译期就存在的闸门，把「拼错的路由名」从线上事故降级成一次解析失败
  • 结构化输出不等于不用校验：网关和降级模型未必严格执行 schema，解析函数的返回类型应该是「待校验」而不是「已经是」

  Key points

  • Free-text routing fails silently: a correct judgement in prose misses your regex and falls through, looking identical to a wrong judgement in the logs
  • Three holes: wording drifts, there is no confidence signal, and invented route names only fail at runtime
  • One declaration used twice: the schema constrains decoding in the request and validates the response, so the two cannot drift
  • An enum is a gate that exists before the call, turning a misspelled route from an incident into a parse failure
  • Structured output does not remove validation — gateways and fallback models may not enforce the schema, so parsing must return an unvalidated value
• 路由不确定或者路由错误时，系统应该怎么兜底？阈值该怎么定？How should the system handle an uncertain or wrong routing decision, and how do you pick the threshold?

  国内高频海外高频深入#routing#fallback#reliability

  分析过程 · 先想清楚再作答

  1. 题眼在「不确定」和「错误」是两件事。多数人只答重试或人工接管，那是把两个问题揉成一个。区分度在于你能不能先给出一条价值判断，再给策略。
  2. 先立论：**路由到错的子 Agent，比路由失败糟糕得多**。失败你至少知道自己失败了，可以追问一句；错了，接手的子 Agent 完全不知道自己接错了活，会用笃定的语气给出一个格式完整的错误答案，用户不会怀疑，会照着去操作。一个自信的错误答案比一句「我没听清」贵一百倍。
  3. 再给可执行的策略，数字要具体：模型给的置信度低于 0.6，或者路由名不在合法名单里，一律落到兜底的 smalltalk，并在 routingReason 里打上 fallback 前缀加原因码（低置信度、未知路由、结构非法各一种）。兜底那位的人设是「信息不足先追问一句缺的关键信息，不要猜」——兜底的本质是把不确定性还给用户。
  4. 阈值怎么定这一问是重点，别背数字：**取决于两类错误哪一类更贵**。客服场景里多问一句只是用户小小的不耐烦，派错可能变成一条错误的退款承诺，所以宁可保守取 0.6；内部工具型 Agent 里多问一句反而更烦人，阈值就该放低。再补一句可落地的定法：拿标准样本集扫一遍，画出不同阈值下的误派率与追问率，选拐点。
  5. 必须点破的一个坑：**置信度是模型自己报的，它不是概率**。模型说 0.9 不代表有九成对。它只是同一模型、同一提示词下相对可用的排序信号，只能当闸门用，不能拿去算期望值。真正的准确率要靠离线评估去量。
  6. 可以预期的追问：兜底会不会把问题掩盖掉？答案是不会，前提是你记了原因码——把一周内落进兜底的请求按原因分组，能直接看出分诊提示词缺了哪一类描述。兜底是止血，原因码才是治本的输入。

  How to reason about it · think before answering

  1. The hinge is that uncertain and wrong are two different failures. Most candidates answer retry or escalate to a human, collapsing both into one. The discriminator is stating a value judgement before giving a policy.
  2. The claim first: routing to the wrong sub-agent is far worse than failing to route. A failure announces itself and lets you ask a clarifying question. A wrong route does not — the receiving agent has no idea it got the wrong job and will produce a confident, well-formatted, wrong answer that the user will act on. A confident wrong answer costs a hundred times more than I did not catch that.
  3. Then give a concrete policy with real numbers: if the model's confidence is below 0.6, or the route name is not in the allowed list, fall back to the small-talk agent and stamp the reason with a fallback prefix plus a cause code (low confidence, unknown route, invalid shape). The fallback agent's job is to ask for the one missing detail rather than guess — falling back means handing the uncertainty back to the user.
  4. The threshold question is the real test, so do not recite a number: it depends on which error is more expensive. In customer support one extra question costs mild annoyance while a misroute can become a wrong refund promise, so stay conservative. For an internal tool the extra question is the bigger cost, so lower it. Then give a method: sweep thresholds over a golden set, plot misroute rate against clarification rate, and pick the knee.
  5. Name the trap: the confidence number is self-reported and is not a probability. Nine tenths does not mean nine in ten are right. It is a usable ranking signal within one model and one prompt — good as a gate, useless for expected-value math. Real accuracy comes from offline evaluation.
  6. Expect: does falling back just hide the problem? Not if you record cause codes. Group a week of fallbacks by cause and you can see exactly which intent the routing prompt fails to describe. The fallback stops the bleeding; the cause code is what fixes it.

  答题要点

  • 先分清两件事：路由失败可以追问，路由错误会让子 Agent 自信地给出错误答案，后者贵得多
  • 策略：置信度低于 0.6 或路由名不在名单里，一律落兜底的 smalltalk，并在 routingReason 打上 fallback 前缀加原因码
  • 兜底不是随便找个人接，而是把不确定性还给用户——兜底那位应当追问缺失的关键信息而不是猜
  • 阈值取决于两类错误哪一类更贵：客服场景多问一句便宜、派错很贵，所以保守；定法是拿标准样本集扫阈值找拐点
  • 置信度是模型自报的，不是概率，只能当闸门用；真正的准确率要靠离线评估量
  • 落兜底时记原因码，按原因分组就能看出分诊提示词缺了哪一类描述

  Key points

  • Separate the two: a failed route can ask a clarifying question, a wrong route produces a confident wrong answer, and the second is far costlier
  • Policy: confidence below 0.6 or a route outside the allowed list falls back to small talk, stamped with a fallback prefix and a cause code
  • Falling back is not picking someone at random — the fallback agent asks for the missing detail instead of guessing
  • The threshold depends on which error costs more; sweep it over a golden set and pick the knee between misroutes and clarifications
  • Self-reported confidence is not a probability — use it as a gate only, and measure real accuracy offline
  • Record cause codes on every fallback; grouping them shows which intent the routing prompt fails to describe

D17 Planner–Executor–Critic + 共享工作区：workspace state、toolBudget、并行 fan-out、review 回路

• 为什么要给每个子任务设 toolBudget 这样的预算？超了预算之后你会怎么处理？Why give each subtask a tool-call budget, and what do you do when it runs out?

  国内高频海外高频进阶#cost-control#reliability#agent-design

  分析过程 · 先想清楚再作答

  1. 这题的题眼在后半句。前半句几乎人人会答「防止成本失控」，真正拉开差距的是超限之后的动作——答「抛异常」的人基本没做过面向用户的 Agent。
  2. 先把「为什么」说具体。子任务卡住的典型形态不是报错，而是反复查、反复不满意、再查——模型不会喊累，它会把额度花光为止。整轮对话的成本封顶是外层的闸，子任务预算是内层的闸；粒度细到单件事的好处是超支时你能精确指出是哪一件失控了，而不是只看到这次对话贵了。
  3. 再点一个容易被忽略的设计点：预算必须是子任务级的，不是单次执行级的。有评审回路时，被打回重做也得计费，否则打回两次实际额度就翻三倍，这道闸等于没设。
  4. 结论是超限的处理：降级返回已有结果并打上标记，不抛错。理由要说透——抛错等于把「这件事只做了一半」升级成「整个请求失败」，用户等了几秒最后看到一句服务异常，可他其实只是没拿到三件事里的一件。给出两件半的答案并说明哪半件没做成，永远比一个错误页有用。
  5. 降级标记本身也要说：它让降级变成可观测、可统计的事实，而不是日志里的一句话。上层据此决定要不要转人工，监控据此画降级率——两个平均分一样的系统，降级率百分之三十和百分之三完全不是一回事。
  6. 可以预期的追问：预算该设多少？答案是从「这件事正常需要几次工具调用」反推再留一点余量，不是拍脑袋取整数；同时要有第二个维度的闸（挂钟时间或 token 数），因为一次超长的工具调用同样能拖垮请求，而它只算一次。

  How to reason about it · think before answering

  1. The hinge is the second half. Everyone can say it controls cost; what separates people is what happens when the budget runs out. Answering throw an exception usually means you have never shipped a user-facing agent.
  2. Make the why concrete: a stuck subtask rarely errors — it queries, dislikes the result, and queries again. The model never gets tired; it will spend whatever you allow. A per-conversation cap is the outer gate, a per-subtask budget is the inner one, and the finer grain tells you which piece went out of control instead of only that the conversation was expensive.
  3. Add the design point people miss: the budget must be per subtask, not per execution. With a review loop, retries have to draw on the same budget, or two rejections triple the real allowance and the gate is meaningless.
  4. The conclusion is the exhaustion path: degrade — return what you already have with a flag — rather than throw. Explain why: throwing upgrades this piece is half done into the whole request failed. The user waited several seconds and gets an error page, when in reality only one of three pieces is missing. Two and a half answers plus a clear note beats an error page every time.
  5. Say something about the flag too: it turns degradation into an observable, countable fact instead of a log line. The layer above decides whether to escalate to a human, and monitoring plots a degradation rate — two systems with the same average score but 30 percent versus 3 percent degradation are not the same system.
  6. Expect: how big should the budget be? Derive it from how many tool calls the task normally needs plus margin, not a round number pulled from the air. And pair it with a second dimension — wall-clock or tokens — because one very slow tool call can ruin a request while counting as a single call.

  答题要点

  • 子任务卡住的典型形态是反复查而不是报错，模型会把额度花光为止；整轮封顶是外层闸，子任务预算是内层闸，细粒度让你能定位到是哪一件失控
  • 预算必须是子任务级而不是单次执行级，否则被评审打回两次实际额度就翻三倍
  • 超限必须降级返回已有结果并标记，不能抛错——抛错把「做了一半」升级成「整个请求失败」，用户连已经查到的部分都拿不到
  • 降级标记让降级率变成可统计指标，上层据此决定转人工，评估据此区分两个平均分相同的系统
  • 预算大小从这件事正常需要几次工具调用反推并留余量，同时配一个时间或 token 维度的闸

  Key points

  • A stuck subtask loops rather than errors, and the model will spend whatever you allow; a conversation cap is the outer gate, a subtask budget the inner one that localises the blowup
  • The budget must be per subtask, not per execution, or two review rejections triple the real allowance
  • On exhaustion, degrade and flag rather than throw — throwing upgrades half done into whole request failed and discards what was already retrieved
  • The degradation flag makes the degradation rate a real metric for escalation and evaluation
  • Size the budget from the task's normal tool-call count plus margin, and pair it with a wall-clock or token gate
• Critic 的评审回路怎么防止陷入死循环？除了次数上限还有什么要防的？How do you keep a Critic review loop from spinning forever, and what else needs guarding besides a retry cap?

  国内高频海外高频进阶#reflection#loop-guard#reliability

  分析过程 · 先想清楚再作答

  1. 问「除了次数上限还有什么」，说明面试官已经预设你会答上限，真正在考的是你有没有真的跑过这条回路。只答上限的人拿基础分，能说出另外两种失效方式的才算过。
  2. 第一种就是无限打回：每改一版评审者挑一个新毛病，永远收敛不了。上限的作用不是省钱，是**保证流程一定会结束**。本课取最多打回 2 次、共 3 次执行，这个量级的取法是「一次有效的修改通常在第二次就完成，第三次还不行说明判据本身有问题」。
  3. 第二种是打回不说人话：评审者只回一句「不合格」，执行者拿不到可执行信息，第二稿原样再交一遍，于是必然打满上限、白烧三倍的钱。所以打回必须带具体理由，而且理由要回写进子任务的目标里带给执行者——「缺了退款结论，请补上」才是可执行的，「质量不佳」不是。
  4. 第三种最危险也最少被提到：评审者和执行者用同一个模型、同一套提示词时，它倾向于认可自己的输出。同一个模型对「什么算好答案」的偏好是一致的，让它复核自己刚写的东西，通过率会高得离谱，这道工序等于没有。缓解手段按性价比排：给评审者一份可核对的客观验收要求；换一个不同的模型来评审，哪怕更便宜；把评审做成逐条打分而不是一句结论。
  5. 还要点一句框架的兜底与业务上限的区别：编排框架通常自带一个递归步数上限，但那是最后一道保险丝，不能当业务上限用——它是全图的，你不知道是哪条回路失控；而且它触发时抛异常，你连已有结果都拿不到，正好违背「降级返回」的原则。
  6. 可以预期的追问：上限用完了返回什么？答：返回已有结果并标记降级，同时把最后一次的评审意见一起带出去，让上层能判断要不要转人工——这条回路的价值不只是修好，还包括「修不好时说清楚差在哪」。

  How to reason about it · think before answering

  1. Asking what else besides a cap tells you the interviewer already expects the cap. What is really being tested is whether you have run this loop for real. The cap earns baseline credit; naming the other two failure modes is what passes.
  2. Failure one is infinite rejection: every revision draws a new complaint and nothing converges. The cap exists to guarantee termination, not to save money. Two rejections and three executions is a reasonable default, because an effective fix usually lands on the second attempt — if the third still fails, the rubric itself is the problem.
  3. Failure two is a rejection with no actionable content. If the reviewer only says not good enough, the executor has nothing to act on and resubmits the same thing, burning the full cap. Rejections must carry a specific reason, and that reason must be written back into the subtask goal. Missing the refund conclusion, please add it is actionable; poor quality is not.
  4. Failure three is the dangerous one people rarely mention: when reviewer and executor share a model and a prompt, the reviewer tends to approve its own output. A single model has consistent preferences about what a good answer looks like, so pass rates go implausibly high and the review step becomes theatre. Mitigations by value: give the reviewer an objective, checkable rubric; use a different model even a cheaper one; score item by item rather than emitting one verdict.
  5. Also distinguish the framework's safety net from your business cap: orchestration frameworks usually ship a recursion limit, but that is a last-resort fuse — it is graph-wide so you cannot tell which loop ran away, and it throws, which means you lose the partial results you were supposed to degrade to.
  6. Expect: what do you return once the cap is used up? Return what you have, flag it as degraded, and carry the last review comment out with it so the layer above can decide whether to escalate. The loop's value is not only fixing things — it is stating precisely what could not be fixed.

  答题要点

  • 次数上限的作用是保证流程一定会结束，不是省钱；本课取最多打回 2 次、共 3 次执行
  • 打回必须带具体、可执行的理由并回写进子任务目标，只说「不合格」会让执行者原样重交、必然打满上限
  • 最危险的是评审者与执行者同模型同提示词，它倾向于认可自己的输出，通过率虚高、这道工序等于没有
  • 缓解手段：给客观可核对的验收要求、换一个模型来评审、逐条打分而不是一句结论
  • 框架自带的递归上限只是保险丝，不能当业务上限：它是全图的、触发时抛异常，连已有结果都拿不到
  • 上限用完要返回已有结果加降级标记，并把最后一次评审意见带出去，供上层决定是否转人工

  Key points

  • A retry cap exists to guarantee termination, not to save money — two rejections, three executions total
  • Rejections must carry specific, actionable reasons written back into the subtask goal; not good enough guarantees an identical resubmission and a maxed-out cap
  • The most dangerous failure is a reviewer sharing model and prompt with the executor: it approves its own output, pass rates inflate, and the step becomes theatre
  • Mitigate with an objective checkable rubric, a different model for review, and item-by-item scoring instead of a single verdict
  • The framework's recursion limit is a fuse, not a business cap: it is graph-wide and it throws, so you lose the partial results you meant to degrade to
  • When the cap is spent, return what you have with a degraded flag plus the last review comment so the layer above can escalate

D18 历史保真与摘要、多模态占位、checkpointer 持久化

• 长对话做摘要时，怎么保证关键信息不丢？在多 Agent 场景下这件事有什么特别的？When summarizing a long conversation, how do you keep the critical information from being lost — and what is different about this in a multi-agent system?

  国内高频海外高频进阶#context-compression#multi-agent#reliability

  分析过程 · 先想清楚再作答

  1. 题眼在后半句。只答「保留用户约束、保留最近几轮」是单 Agent 的标准答案，能过但不出彩；面试官问「多 Agent 有什么特别的」，是在看你有没有真的在协作图里踩过这个坑。
  2. 先把单 Agent 那半答扎实：触发用阈值不用定时器，本课口径是消息超过 20 条或估算超过 8000 token（一个字符算一个 token，故意高估，低估会让阈值永远触发不了）；保留最近 6 条原文不动；切口必须对齐到一轮的开头，切在工具调用与工具结果之间会让下一次请求出现悬空消息，多数厂商直接返回 400。
  3. 然后给出多 Agent 那半的关键差别：单 Agent 里摘要丢的是**细节**，多 Agent 里摘要丢的是**判据**。评审者判一份产出合不合格，靠的是分清「这句是待验收的产出、那句是验收要求」；一段把发言人抹平的流水摘要读起来通顺，但评审拿它做不了任何判断。
  4. 所以多 Agent 的摘要有一条额外硬要求：每条被压掉的消息，在摘要里都要留下「第几条 + 谁说的」这两个坐标。实现只有一行——把转录写成带序号和角色前缀的形式再交给模型。
  5. 再补一条边界，这条最能显出你写过：**摘要只对自然语言历史动手，不碰任何结构化字段**。把共享工作区压成一句话，「按 id 找到某条子任务、比对验收要求」就整个失效了，结构化数据压成自然语言就再也回不去。附件字段更是碰不得——它存的是引用不是内容，摘要掉等于把那个对象变成孤儿。
  6. 可以预期的追问：摘要用哪个模型、失败了怎么办？答可以用更便宜的小模型（它只做归纳不做推理），失败时的正确行为是**跳过这一轮压缩继续跑**并告警，而不是让整次执行失败——阈值定在七八成就是为了留出这次抢救余量。

  How to reason about it · think before answering

  1. The hinge is the second half. Answering only keep user constraints and the last few turns is the standard single-agent answer — passable, not memorable. Asking what is different in multi-agent is asking whether you have actually hit this in a collaboration graph.
  2. Get the single-agent half solid first: trigger on thresholds, never a timer. This course uses more than 20 messages or an estimated 8000 tokens, counting one character as one token — deliberately high, because underestimating means the threshold never fires. Keep the last 6 messages verbatim. Align the cut to a turn boundary: cutting between a tool call and its result produces a dangling message and most vendors return 400.
  3. Then name the real difference: in a single agent a summary loses detail; in a multi-agent graph a summary loses the criteria. A critic decides whether output passes by telling apart what is being reviewed from what the requirement was. A smooth narrative summary that flattens speakers reads fine and is useless to the critic.
  4. So multi-agent summarization has one extra hard requirement: every compressed message must leave behind two coordinates — its index and its speaker. The implementation is one line: build the transcript with numbered, role-prefixed entries before handing it to the model.
  5. Add the boundary that shows you have shipped this: summarize natural-language history only, never structured fields. Compressing the shared workspace into a sentence kills every lookup by task id and every comparison against an acceptance requirement, and structured data does not come back. Attachments are even more off-limits — they hold a reference, not content, so summarizing one orphans the underlying object.
  6. Expect the follow-up: which model writes the summary, and what if it fails? A cheaper small model is fine since the job is condensation, not reasoning. On failure the correct behaviour is to skip this round of compression, keep running, and alert — not to fail the whole execution. Setting the threshold at seventy or eighty percent exists precisely to leave that rescue room.

  答题要点

  • 触发用阈值不用定时器：超过 20 条或估算超过 8000 token，token 按一字符一 token 保守高估
  • 保留最近 6 条原文不动，切口必须对齐到一轮开头，否则会出现有调用没结果的悬空消息、请求直接 400
  • 多 Agent 的差别：摘要丢的不是细节而是判据，评审者靠「谁在第几步说的」区分产出与验收要求
  • 所以每条被压掉的消息都要在摘要里留下条号与发言人，实现就是把转录写成带序号和角色的形式
  • 只压自然语言历史，不碰共享工作区这类结构化字段，更不能碰存引用的附件字段
  • 摘要可用更便宜的小模型；摘要调用失败时跳过这一轮压缩并告警，不要让整次执行失败

  Key points

  • Trigger on thresholds, not timers: more than 20 messages or an estimated 8000 tokens, counting one character as one token to stay conservative
  • Keep the last 6 messages verbatim and align the cut to a turn boundary, or you ship a dangling tool call and the request 400s
  • The multi-agent difference: a summary loses criteria, not just detail — the critic needs to know who said what and at which step
  • So every compressed message keeps its index and speaker in the summary; the implementation is a numbered, role-prefixed transcript
  • Summarize natural-language history only — never the shared workspace or other structured fields, and never the attachment references
  • A cheaper small model is fine for summarizing; if the call fails, skip compression for this round and alert rather than failing the run
• 从 checkpoint 恢复执行（replay）需要注意什么？说几个真实会踩的坑。What do you need to watch out for when replaying execution from a checkpoint? Give failure modes you would actually hit.

  国内高频海外高频深入#checkpointing#replay#reliability

  分析过程 · 先想清楚再作答

  1. 这题最容易答成「读出来接着跑就行」。区分度在于你能不能说出**这些坑几乎全是静默的**——不抛异常、日志干净、结果看起来也对，只有对比数据时才发现不对。能说出这一点，答案就已经赢了一半。
  2. 先给一条推导链：检查点里存的是「当时那个版本的代码眼里的状态形状」，恢复就是把它塞回今天这个版本的代码里。所以所有坑都来自**两端不一致**：数据的形状、执行的入口、和那些不该被重放的东西。
  3. 坑一，恢复时又把输入喂了一遍。恢复的入口是不带输入地调用，状态已经在检查点里；带着原来那句话再调一次，框架会把它当成一次新的状态更新叠在中断点上，历史变成两份。它不报错。
  4. 坑二，分叉忘了带检查点 id。只给会话 id 拿到的是这条线最新的状态，于是「从第 2 步重来」变成了「在最后一步后面接着写」。同样不报错，只有对比子任务列表才看得出来。
  5. 坑三，版本兼容。改一个字段名、加一个必填字段，库里的老检查点就和新代码对不上；而缺字段读出来是 undefined，拼进文案就是字符串「undefined」，参与算术就是 NaN——比如工具预算的上限判断，一旦变成 NaN 比较，恒为假，预算上限在恢复出来的那条线上彻底失效。正确做法是在读的那一侧迁移，迁移函数只补默认值和改名、不做业务判断，绝不能失败。
  6. 坑四，不该被重放的东西进了状态。一次性的人工干预（比如人工改派）如果写进图状态，就会被检查点持久化并在每次恢复时重放一遍。判断口径：这条信息说的是「这一次执行怎么跑」还是「这个会话是什么」，前者进运行时配置，后者才进状态。
  7. 可以预期的追问：待执行的并行子任务存不存？答存——检查点里除了状态快照还有一份「还没跑的那几步，连参数一起」，所以恢复不用重跑规划节点；但它存在框架的内部通道里，自研存储层只实现「存状态」而漏掉这一半，恢复出来的图会看起来跑完了、其实一件活都没派出去。

  How to reason about it · think before answering

  1. The easy failure is answering just load it and keep going. The discriminator is recognising that almost every replay bug is silent — no exception, clean logs, plausible output, and you only notice when you diff the data. Saying that up front wins half the question.
  2. Give a chain first: a checkpoint stores the state shape as the code of that moment understood it, and replay pushes it back into today's code. So every failure comes from a mismatch across those two ends — the shape of the data, the entry point of execution, and things that should never have been replayed at all.
  3. Trap one: feeding the input again on resume. Resume takes no input; the state is already in the checkpoint. Passing the original message once more makes the framework treat it as a fresh update stacked on the interrupt point, and the history quietly doubles. Nothing throws.
  4. Trap two: forking without a checkpoint id. With only the thread id you get that thread's latest state, so start over from step 2 silently becomes append after the last step. Again nothing throws; you only see it by diffing the task list.
  5. Trap three: version drift. Rename a field or add a required one and every old checkpoint stops matching the new code. A missing field reads as undefined, which renders as the literal string undefined in user-facing text and as NaN in arithmetic — a tool-budget ceiling compared against NaN is always false, so the budget silently stops existing on resumed threads. Migrate on read, and keep the migration to defaults and renames only: it must never fail.
  6. Trap four: replayable data that should not be replayed. A one-off human override written into graph state gets checkpointed and re-applied on every resume. The test: does this describe how this run executes, or what this conversation is? The former belongs in runtime config, only the latter in state.
  7. Expect the follow-up: are pending parallel tasks preserved? Yes — a checkpoint holds not just the state snapshot but the steps not yet run, arguments included, so the planner does not re-run. But they live in a framework-internal channel, so a hand-rolled store that persists state and forgets that half will resume into a graph that looks finished while no work was ever dispatched.

  答题要点

  • 先点破共性：replay 的坑几乎全是静默的，不报错、日志干净、结果看着也对
  • 恢复不要带输入，带了就是在中断点上又追加一次，历史变成两份
  • 分叉必须带检查点 id，只给会话 id 会落在最新状态上，「从第 2 步重来」变成「接着往后写」
  • 版本兼容：缺字段读出来是 undefined 或 NaN，会让预算上限之类的比较恒为假；在读的那一侧迁移，迁移只补默认值和改名且不能失败
  • 一次性的人工干预不要进图状态，否则会被持久化并在每次恢复时重放；「这次怎么跑」进配置，「这个会话是什么」才进状态
  • 待执行的并行子任务连参数一起存在检查点里，所以恢复不重跑规划；自研存储层漏掉这一半，恢复出来的图会一件活都不派

  Key points

  • Lead with the pattern: replay bugs are almost all silent — no exception, clean logs, plausible output
  • Resume takes no input; passing one appends another update at the interrupt point and doubles the history
  • Forking requires the checkpoint id — thread id alone lands on the latest state, turning start over from step 2 into append after the end
  • Version drift: missing fields read as undefined or NaN, so comparisons like a tool-budget ceiling become permanently false. Migrate on read, restricted to defaults and renames, and never let it fail
  • Keep one-off human overrides out of graph state or they get persisted and re-applied on every resume — how this run executes belongs in config, what this conversation is belongs in state
  • Pending parallel tasks are stored with their arguments, so the planner does not re-run; a hand-rolled store that skips that half resumes into a graph that dispatches nothing

D21 评估与可观测：golden set、LLM-as-judge、tracing、失败率/成本面板；Pi vs LangGraph 总结；W3 复盘

• 用大模型给大模型的输出打分（LLM-as-judge），有哪些不可靠的地方？怎么办？What makes LLM-as-judge unreliable, and what do you do about it?

  国内高频海外高频深入#evaluation#llm-as-judge#reliability

  分析过程 · 先想清楚再作答

  1. 这题筛的是「你是真用过，还是听说过」。用过的人能报出具体的失效形态和量级，没用过的人只会说「可能不准」。
  2. 第一种，**同源偏差**：judge 和被评估的 Agent 用同一个模型时，它偏向认可自己的输出——同一个模型对「什么算好答案」的偏好是一致的，让它复核自己刚写的东西，它当然觉得没问题。实测数量级：同一批被改坏的产出，同源 judge 给 14/15，换个模型只给 12/15，被多放过去的正是最该抓的边缘产出。这个坑不止在 judge，**凡是「模型评模型」的位置都有**，Critic 节点是同一个问题。
  3. 第二种，**长度偏好**：judge 倾向给篇幅大的答案更高分。实测：一条 33 字的正确回复灌上 141 字无关客套话，凭印象打分的提示词就从 2 分涨到 4 分，内容一个字没变。
  4. 第三种，**评分提示词漂移**：judge 的评分随提示词微调整体移动。同一份产出，两套评分提示词一套给 2 分一套给 5 分。所以有条硬纪律——**分数只在同一套 judge 提示词内部可比**，跨版本比较是没有意义的。
  5. 解药要一一对应，别笼统说「多测几次」：固定 judge 提示词并版本化（每条评分记录带上 rubric 版本与 judge 模型，那是它的坐标；面板发现混了两套口径应当直接拒绝聚合，而不是算出一个没含义的平均分）；默认用不同的模型当 judge，而且这该是默认值不是可选项；留一小批人工标注做校准集，每次改评分提示词拿它对一遍，**比的是结论（过或不过）而不是分数差**——差 1 分无所谓，结论翻了就是事故。
  6. 还有一条更根本的：**把评分标准从主观印象换成可核对的清单**，它同时解掉长度偏好——照清单逐条数，灌水加不了分。实测那条灌水回复在清单口径下前后都是 5 分，纹丝不动。
  7. 可以预期的追问：judge 便宜还是人工便宜？答：judge 的成本和被评估的系统本身一个量级，所以「跑一次全量评估多少钱」是你决定每次提交都跑还是每天跑一次的依据；而人工的成本不在钱在延迟——它给不了你改一次提示词就想看一次结果的反馈速度，所以人工只该用在校准集上。

  How to reason about it · think before answering

  1. This screens for whether you have actually used it. People who have can name specific failure shapes with magnitudes; people who have not just say it might be inaccurate.
  2. First, self-preference: when the judge and the evaluated agent share a model, it favours its own output — the same model has a consistent notion of what a good answer looks like, so asking it to review what it just wrote gets an approving verdict. Measured: on the same batch of deliberately degraded outputs, a same-model judge gave 14/15 while a different model gave 12/15, and the extra passes were exactly the borderline cases worth catching. This is not confined to judges — every model-grading-model position has it, and a Critic node is the same problem.
  3. Second, length bias: judges reward longer answers. Measured: padding a correct 33-character reply with 141 characters of irrelevant pleasantries moved an impression-based rubric from 2 to 4 without changing a word of substance.
  4. Third, rubric drift: scores shift wholesale when the judge prompt is tweaked. The same output scored 2 under one rubric and 5 under another. Hence the hard rule: scores are comparable only within one judge prompt, and cross-version comparison is meaningless.
  5. Match each remedy to its failure rather than saying run it a few more times. Freeze and version the judge prompt — every score record carries its rubric version and judge model, which are its coordinates, and a dashboard that finds two rubrics mixed should refuse to aggregate rather than emit a meaningless average. Default to a different model as judge, as a default and not an option. Keep a small human-labelled calibration set and re-run it whenever the rubric changes, comparing verdicts (pass or fail) rather than score deltas — one point of drift is fine, a flipped verdict is an incident.
  6. And one deeper fix: replace impressionistic criteria with a checkable list, which also dissolves length bias — counting items off a list gives padding nothing to earn. Measured, that padded reply scored 5 both before and after under the checklist rubric.
  7. Expect: is a judge cheaper than humans? The judge's cost is the same order as the system being evaluated, so what a full evaluation run costs decides whether you run it per commit or nightly. Human cost is not money but latency — it cannot give you feedback at the speed of one prompt edit, which is why humans belong on the calibration set only.

  答题要点

  • 同源偏差：judge 与被评估 Agent 同模型会虚高（实测 14/15 vs 异源 12/15），且凡「模型评模型」的位置都有，Critic 同理
  • 长度偏好：灌水 141 字能让印象分从 2 涨到 4，内容一字未变
  • 评分提示词漂移：同一产出两套 rubric 一个 2 分一个 5 分，所以分数只在同一套提示词内部可比
  • 解药一一对应：rubric 版本化并随记录存坐标、面板发现混口径直接拒绝聚合、默认换模型当 judge
  • 留人工标注校准集，比结论（过/不过）而不是比分数差——差 1 分无所谓，结论翻了是事故
  • 更根本的是把主观印象换成可核对的清单，同时解掉长度偏好（清单口径下灌水前后都是 5 分）

  Key points

  • Self-preference: a same-model judge inflates scores (14/15 vs 12/15 cross-model), and it applies to every model-grading-model spot including Critic
  • Length bias: 141 characters of padding moved an impression score from 2 to 4 with no substantive change
  • Rubric drift: the same output scored 2 and 5 under two rubrics, so scores compare only within one judge prompt
  • Remedies map one-to-one: version the rubric and store it alongside each record, refuse to aggregate mixed rubrics, default to a different judge model
  • Keep a human-labelled calibration set and compare verdicts, not score deltas — a point of drift is fine, a flipped verdict is an incident
  • The deeper fix is a checkable list instead of impressions, which also removes length bias (the padded reply scored 5 both ways)