把"全自治运营"拆开看，里面是二十多个 LLM 驱动的 Agent：有人负责回复，有人只负责挑错，有人记住每段关系，有人维护知识库，有人把运营者的一句话编排成批量动作。它们认知隔离、各管一段，再由确定性的调度器自主串起来。
Open up "autonomous operations" and you'll find twenty-plus LLM-driven agents: one replies, one only catches mistakes, one remembers every relationship, one tends the knowledge base, one turns a single operator sentence into a batch of actions. They're cognitively isolated, each owns one slice, and a deterministic scheduler wires them together.
客户发来一句话，先由回复决策 Agent 拿主意，再交给一个看不到它内心戏的独立评审 Agent 打分挑错——写回复的和挑错的，从来不是同一个大脑。
When a message arrives, the reply agent decides; then an independent reviewer that can't see its inner reasoning scores and catches errors — the one writing and the one checking are never the same brain.
决策、评审、建画像——驱动每一次自动回复的三个 Agent。
Decide, review, profile — the three agents behind every auto-reply.
系统的"主嘴"。读完上下文、记忆、知识与画像后，决定这一轮说什么、推进到哪一步、要不要发素材或引荐名片。
The system's voice. After reading context, memory, knowledge and the profile, it decides what to say this turn, how far to advance, whether to send assets or refer a card.
专职挑错。只拿到事实面投影，看不到回复 Agent 的自我推理，独立给五个维度打分——事实风险、压力感、拟人度、情感价值、产品准确性。
The dedicated critic. Gets only a facts projection — never the reply agent's reasoning — and scores five dimensions independently: fact risk, pressure, human-likeness, emotional value, product accuracy.
第一次接触某位客户时建立基础画像：从已有备注与开场对话里，给出初步的身份、阶段与关系判断，作为后续经营的起点。
On first contact, it builds a base profile: from existing notes and the opening exchange it drafts an initial identity, stage and relationship read — the starting point for everything after.
跨轮、长期。回复发出之后，这两个 Agent 负责"记住这个人"——读懂反应、沉淀记忆，让下一次对话接得上。它们产出的标签与画像并非都可信：哪些进决策、哪些只做记录，由信息信任层级严格分层。
Cross-turn and long-lived. After a reply goes out, these two agents "remember the person" — reading reactions and consolidating memory so the next turn picks up where it left off. Not all the tags and profiles they produce are trusted equally: what drives decisions versus what's record-only is strictly split by the information trust tiers.
回复发出后，分析客户的下一句话是什么态度——认可、犹豫、抗拒还是冷场，把"刚才那句说得对不对"变成可沉淀的信号，反哺下一轮决策。
After a reply, it reads the customer's next message — agreement, hesitation, resistance or silence — turning "did that land?" into a signal that feeds back into the next decision.
把零散的对话压缩成结构化的长期记忆卡：确认的事实、关系状态、产品契合、下一步动作。过期事实会被显式标记，不再被引用。
Compresses scattered dialogue into a structured long-term memory card: confirmed facts, relationship state, product fit, next action. Stale facts are explicitly deprecated and no longer cited.
知识从哪来、怎么查、缺了谁来补、过期谁来修——一整套围绕知识库的 Agent，唯独不碰"自动判定为真"那条红线（新知识永远先落草稿待人核验）。
Where knowledge comes from, how it's retrieved, who fills gaps, who repairs staleness — a full set of agents around the knowledge base, never crossing the one red line: it never auto-verifies (new knowledge always lands as a draft awaiting human review).
渐进式检索的"规划大脑"：看目录 → 下钻文档 → 展开切片，多轮工具调用按需取证，token 级流式吐答案。只读，绝不改写知识。
The planning brain for progressive retrieval: catalog → drill into a doc → open the slice, fetching evidence on demand over tool-calling rounds, streaming the answer token by token. Read-only, never rewrites knowledge.
查无可引用知识时，不编造——而是生成一句面向运营者的精炼追问，引导把缺失的知识补进库里。把"答不上来"变成"知道该补什么"。
When there's nothing to cite, it doesn't fabricate — it generates one crisp question for the operator, guiding them to fill the missing knowledge. Turns "can't answer" into "knows what to add".
运营者用大白话就能补库。意图识别 → 起草切片 → 更新 → 澄清四个 prompt 协同，把一段对话变成结构化知识草稿。
Operators grow the base in plain language. Four prompts work together — intent, draft, update, clarify — turning a conversation into a structured knowledge draft.
切片过期、锚点失效或内容存疑时，提出修复方案——只是"提案"，落地与否仍需人确认，守住"AI 不自证为真"。
When a chunk goes stale, an anchor breaks or content is doubtful, it proposes a fix — only a proposal; whether it lands still needs human confirmation.
辅助核验，而非自动放行。给出一致性与可信度评估供人参考，但"判定为真"的最终权永远在人手里——这是不可逾越的红线。
It assists verification, never auto-approves. It offers a consistency and trust assessment for humans, but the final "this is true" call always stays with a person — a hard line.
导入 PDF、图片等素材时先生成结构化预览，并自动抽取标签。无论哪种入口，新知识一律落草稿、标记待核验。
When importing PDFs, images and more, it generates a structured preview and auto-extracts tags. Whatever the entrypoint, new knowledge always lands as a draft, marked for review.
两个 Agent 接力：一个分析当天知识使用日志、识别健康信号，一个把分析成文为日报卡。让知识库的"用得好不好"每天看得见。
Two agents in relay: one analyzes the day's knowledge-usage logs and spots health signals, the other composes them into a digest card — making "how well knowledge performs" visible daily.
上线一个新行业不用写代码。这组 Agent 帮运营者把"我想做什么生意"草拟成可运行的运营方法、行业画像与状态机。
Onboarding a new industry takes no code. These agents help operators draft "what business I run" into runnable playbooks, domain profiles and state machines.
从运营者的业务描述生成一套完整的运营方法论（playbook），并能基于运行表现持续优化。这是"全自治"能落到任意行业的关键一环。
Generates a complete playbook from the operator's business description, and refines it from runtime performance. The key to landing "autonomous" in any industry.
引导向导里，AI 帮你起草这个行业该有的客户阶段、标签维度、状态机与人格基调，并实时预览效果。新行业从"零配置"到"能跑"，由 AI 陪着搭。
In the setup wizard, the AI drafts the customer stages, tag dimensions, state machine and persona this industry needs, with a live preview. From zero config to runnable, with the AI alongside.
自治不只在客户侧。运营者一句话就能驱动批量动作，改 prompt 要先过红线裁判，遇到超出职权的事 AI 会向幕后决策源请示——拿回结论后仍用自己的口吻向客户转述。
Autonomy isn't only customer-side. One operator sentence drives a batch of actions, prompt edits face a red-line judge first, and on matters beyond its authority the AI consults the behind-the-scenes decision source — then relays the conclusion to the customer in its own voice.
运营者用一句自然语言（"给所有犹豫期客户发个关怀"），它就编排成一份带风险分级的操作计划——观测、调参、改策略、灰度发布、知识维护都能调度。危险与核验类动作恒需人确认。
One operator sentence ("send a check-in to all hesitating customers") becomes a risk-graded action plan — observation, tuning, strategy edits, staged rollout, knowledge upkeep. Dangerous and verify actions always need confirmation.
改提示词的"第三道闸"。语义级判断这次改动有没有变相引入真人接管、削弱知识 grounding，三态裁决：放行 / 拒绝 / 需人确认。裁判掉线则降级为人确认。
The third gate on prompt edits. It judges, at the semantic level, whether a change sneaks in human takeover or weakens grounding — three verdicts: pass / reject / needs confirmation. If the judge is down, it falls back to human confirmation.
遇到超出 AI 职权的事项（如能否破例、特批价格），向幕后决策源（领导）请示，拿回结论后解读成可执行的对客转述。客户始终只跟 AI 对话，从不面对真人。
On matters beyond its authority (an exception, a special price), it consults the behind-the-scenes decision source, then interprets the verdict into an actionable relay. The customer only ever talks to the AI, never a human.
系统会自己提议怎么把话说得更好——但每一条改动都要先在影子里证明，再灰度放量，与生产链路完全隔离。
The system proposes how to speak better — but every change is proven in the shadow first, then rolled out gradually, fully isolated from the production path.
分析失败样本，提出对回复 Agent 提示词的改写方案。它刻意不走统一 LLM 入口、与生产 Reply 链路完全分离，token 单独计入演化预算——演化绝不污染线上。
It analyzes failed samples and proposes rewrites to the reply agent's prompt. It deliberately bypasses the shared LLM entrypoint, stays separate from the production reply path, and bills tokens to a dedicated evolution budget — evolution never contaminates production.
批判 Agent 的每条提案都不会直接上线：先在影子回放里用历史样本验证，达标才进入灰度发布、显著性检验与自动放量/回滚。这套调度是确定性的统计逻辑，不是 LLM。
No proposal goes live directly: it's first proven on historical samples in shadow replay, and only then enters staged release, significance testing and auto rollout/rollback. This orchestration is deterministic statistics — not an LLM.
很多人以为做多行业要给每个行业养一支专属 Agent。我们没有。教育、医美、电商、咨询……用的是同一个回复决策 Agent，差异全部来自注入的行业画像数据——客户阶段、标签维度、状态机、运营方法、人格基调都是配置，不是代码。引擎层与行业彻底解耦：上一个新行业，配的是数据，不是新写一个智能体。这也是为什么"配置生成编队"如此关键——它让通用引擎一键适配任意私域生意。
Many assume multi-industry means a dedicated agent per industry. We don't. Education, aesthetics, e-commerce, consulting — all run on one reply decision agent; the differences come entirely from injected domain-profile data — stages, tag dimensions, state machine, playbook, persona are config, not code. The engine is fully decoupled from any industry: onboarding a new one means configuring data, not writing a new agent. That's why the configuration squad matters — it lets one universal engine adapt to any private-domain business.
我们不把调度也包装成"智能体"。决定何时主动跟进、何时派发、何时唤醒沉默客户的，是可预测、可审计的规则引擎。它们自主发起，但绝不直接说话——真正开口的永远是上面那些 LLM Agent，经过完整的决策与评审链路。
We don't dress up scheduling as "agents" either. What decides when to follow up, when to dispatch, when to re-engage a quiet customer is a predictable, auditable rule engine. They initiate autonomously but never speak directly — the actual talking always goes through the LLM agents above, via the full decision-and-review path.
周期扫描已纳管客户，三段规则识别该不该主动跟进：太久没说话（沉默唤醒）、答应过的事没兑现（承诺兑现）、卡在某阶段不动（阶段停滞）。命中只生成跟进任务，从不直接发消息。
Periodically scans managed customers; three rules decide whether to follow up: gone quiet (re-engage), an unkept promise (commitment), stuck at a stage (stagnation). A hit only emits a follow-up task — it never sends directly.
silent · commitment · stage_stagnation定时轮询到期的跟进任务，逐个送进统一发送网关——和客户主动发消息触发的，是同一条决策+评审+发送流水线。没有"绕过网关"的捷径。
Polls due follow-up tasks on an interval and feeds each into the unified gateway — the same decide-review-send pipeline triggered by an inbound message. No shortcut bypasses the gateway.
默认开 · 复用网关独立 worker 原子抢占待发条目（租约 + 幂等键），调 MCP 工具真正发出。超时由回收器收回租约重试，绝不重复发送。发送与决策解耦，互不阻塞。
An independent worker atomically claims pending entries (lease + idempotency key) and calls the MCP tool to actually send. Expired leases are reclaimed and retried — never double-sent. Sending and deciding are decoupled.
幂等 · 租约回收把上面的编队串成一条线：从客户发来一句话，到 AI 的回复真正送达，中间哪些是确定性调度、哪些是 LLM Agent 在思考，一目了然。
Stringing the squads into one line: from the customer's message to the AI's reply landing, you can see exactly which steps are deterministic scheduling and which are LLM agents thinking.
webhook 接收、落库，校验纳管状态、冷却、频率、每日上限。未纳管的只存不回。或由战略规划器/任务 worker 主动发起。
The webhook receives and persists, checks managed status, cooldown, rate and daily cap. Unmanaged contacts are stored, not answered. Or initiated by the planner / task worker.
永远前置：先决定要不要查知识、查什么、引用哪些切片。查不到就如实标记，绝不假装有据。
Always first: decide whether and what to consult, and which slices to cite. If nothing's found it says so — never fakes grounding.
读完上下文、记忆、知识、画像，渐进式三档按需加载，生成回复草案与风险自评。首次接触会先由初始画像 Agent 建画像。
After context, memory, knowledge and profile — loaded by progressive tier on demand — it drafts a reply with self-assessed risk. First contact starts with the initial-profile agent.
只看事实投影、看不到内心戏，五维打分。要求改写时，回复 Agent 在完整档重生成一次并复审一次——至多一次，不无限循环。
Sees only the facts projection, not the reasoning, and scores five dimensions. If it asks for a rewrite, the reply agent regenerates once on the full tier and re-reviews once — at most once.
三道闸门、状态机校验、发出前再查一次前置条件，approved 才入幂等发件箱，由派发器调 MCP 真正送达。全程留痕、可审计。
Three gates, state-machine validation, a final precheck; only approved enters the idempotent outbox, where the dispatcher calls MCP to deliver. Fully logged and auditable.
送达之后，关系经营编队接手：读懂客户的下一句反应、把这一轮沉淀进长期记忆卡，让下一次对话从这里继续。
After delivery, the relationship squad takes over: read the customer's next reaction and fold this turn into the long-term memory card, so the next conversation continues from here.
以上都是生产链路的 Agent。此外还有一组只在 CI 里跑的质量保障 Agent——异族大模型扮演难缠客户、多裁判交叉打分，专门给运营对话挑刺，从不参与线上对话。看真实大模型测试 →
All of the above are production agents. Beyond them is a set of quality-assurance agents that run only in CI — cross-family models role-playing tough customers, multiple judges scoring in crossfire to stress-test operational dialogue, never part of live conversations. See real-LLM testing →
不是营销话术堆出来的"智能"。每个 Agent 的职责、输入、输出、守门规则，都写进了 Rust 代码，跑在统一的 LLM 入口下，全程留痕、可审计。
This isn't "intelligence" stacked from marketing copy. Each agent's role, inputs, outputs and guardrails are written into Rust, running under one LLM entrypoint — fully logged and auditable.