几千个好友，一个人忙得过来吗？谁买过什么、答应发的资料、上次聊到哪——全靠脑子记、靠翻聊天记录；员工一走，客户和这些记忆一起被带走。WeAgent 用一个自治 Agent 替你经营每一个客户：读懂来意、查知识取证、自己拟回复、独立评审把关、到点主动跟进，全程自治。客户和记忆都沉淀在企业，永远属于公司，不属于某个销售。
Thousands of contacts — can one person keep up? Who bought what, the doc you promised, where the last chat left off — all carried in someone's head and chat history; when an employee leaves, the customers and that memory walk out too. WeAgent runs every customer through one autonomous Agent: reads intent, grounds claims in verified knowledge, drafts the reply, passes an independent review, and follows up on time — autonomous end to end. Customers and memory stay with the company, never with one salesperson.
一个人记不住几千个客户各自的脾气和分寸，系统可以。这些都是我们在设计时抠到很细的地方——而且始终守一条底线：只观测、只沉淀，绝不替你武断操控。它对每个客户的判断和依据，运营者都看得到，方便随时复盘管理。
One person can't hold the temperament and nuance of thousands of customers — the system can. These are the details we sweated over in design, all under one principle: observe and accumulate, never manipulate on a hunch. Every judgment it makes about a customer, and the grounds for it, stays visible to operators for review.
老员工怎么应对、产品怎么介绍，沉淀成可复用的知识库。AI 开口必须有据可查——没核实过的产品说法，宁可不说也绝不瞎编。
How your best reps respond and pitch becomes a reusable knowledge base. Every product claim must be grounded — if it isn't verified, the AI stays silent rather than make it up.
有据才说 · 未核实即拦截Grounded only · unverified is blocked谁怕被打扰、谁喜欢直接报价、上次聊到哪、答应过什么，都记着、还在更新。人维护不了的细致，规模化地维护起来。
Who dislikes being pinged, who wants a price up front, where the last chat ended, what was promised — all remembered and kept current. Nuance no person could hold, maintained at scale.
记得住 · 还在更新Remembered · always current对客户的判断不是一锤定音。每一次新对话都是新证据，系统持续修正认知——越聊越准，而且始终诚实标注「有多确定」。
Its read on a customer is never one-and-done. Each new chat is fresh evidence; the system keeps updating — sharper over time, always honest about how confident it is.
持续修正 · 诚实标注置信Always revising · confidence shown它用心理学公认的「大五人格」(OCEAN) 五个维度慢慢摸清每个客户的性格倾向，沉淀成画像供你参考——只观测、不臆断，证据不足就如实标注「还不确定」。真正因人调整的是关系：对客户、同行、朋友区别对待，口吻和跟进节奏各不相同，不是一套说辞对所有人。
It reads each customer's disposition along the psychology-standard Big Five (OCEAN), kept as a profile for your reference — observed, never presumed; thin evidence is honestly marked "not sure yet." What actually adapts is the relationship: customers, peers and friends are treated differently — tone and cadence shift, never one script for everyone.
大五画像供参考 · 关系分化Big Five profile for reference · per-relationship私域运营真正的难，不在发消息，在「人扛不住规模」：记不住、忙不过来、会离职。每一条，WeAgent 都有对应的解法。
The hard part of private-domain ops isn't sending messages — it's that people can't carry the scale: they forget, get overloaded, and leave. WeAgent answers every one.
很多 AI 客服把人机协作做成「兜不住就转人工」。WeAgent 的红线相反：客户永远只跟 AI 对话、永不直接面对真人。这条红线由代码、CI 静态扫描和发送网关三重守护。
Most AI agents fall back to "escalate to a human." WeAgent does the opposite: the customer only ever talks to the AI. This line is guarded three ways — by code, by a CI lint, and by the send gateway itself.
对话始终是 AI 在说。即使 AI 需要请示领导，客户看到的也只是 AI 自然的回复，从不被丢给真人或晾在一边。
Every message comes from the AI. Even when it consults a decision-maker, the customer only sees a natural AI reply — never a handoff, never left waiting.
遇到超出职权的事，AI 向幕后领导请示、拿回结论，再用自己的口吻向客户转述。领导隐于 AI 之后，客户从不直接对接。
When a decision exceeds its mandate, the AI consults the principal, gets a verdict, and relays it in its own voice. The principal stays behind the AI — invisible to the customer.
账号可显式开启「辅助模式」，让 AI 在高价值时刻引荐真人顾问名片。默认关闭、需审核、发送前二次校验——全自治红线本体一字不动。
An account may opt into "assist mode" to let the AI introduce a human advisor's card at high-value moments. Off by default, review-gated, double-checked before send — the autonomy line stays intact.
从一条入站消息，到决策、评审、发送、复盘、记忆、知识沉淀、自我演化——每一环都落地为真实代码与可审计日志。
From a single inbound message to decision, review, send, retrospective, memory, knowledge curation, and self-evolution — every step is real code with auditable logs.
好友池、单人画像、长期记忆、运营大脑、方法论、模拟验证、会话回放。智能模式与传统模式双视角。
Contact pool, per-person profile, long-term memory, operating brain, playbooks, simulation, conversation replay — smart and classic modes.
9 类知识切片，写入即被编织进互联知识仓。渐进式召回（目录→切片→取证），AI 永不自动核验。
9 wiki types woven into a linked knowledge store on write. Progressive recall (catalog → slice → cite); the AI never self-verifies.
用自然语言操作整个系统：查询、配置、发消息、建任务。执行计划 + 工具调用状态 + Dry-run 演练 + 高风险确认。
Run the whole system in natural language: query, configure, send, schedule. Execution plans, tool-call states, dry-run, and high-risk confirmations.
后台 worker 在影子环境重放历史决策，微调阈值与提示词；显著性达标才可发布，一键回滚。主链路零副作用。
A worker replays past decisions in shadow mode to tune thresholds and prompts; release only on significance, rollback in one click. Zero side effects on the live path.
8 类待办汇于一处：请示裁决、知识核验、标签候选、关系建议、画像发布、进化发布、经验晋升、知识缺口。
8 review streams in one place: escalations, knowledge verification, tag candidates, relationship hints, profile publish, evolution release, lesson promotion, knowledge gaps.
决策日志、MCP 调用、发送账本、任务调度、自治回路监控、发送成效与运营成效全景。
Decision logs, MCP calls, send ledger, task scheduling, autonomy-loop monitoring, plus send and ops outcome analytics.
传统私域工具是「人发指令、机器执行」。WeAgent 是一个真正的自治体：自己决定何时该说话、说什么、要不要请示、发完怎么复盘——四种自治能力各自落地为真实代码与可审计回路，不是 PPT 上的形容词。
Legacy private-domain tools are "human commands, machine executes." WeAgent is a genuine autonomous system: it decides when to speak, what to say, whether to consult, and how to reflect afterward — four autonomy capabilities, each grounded in real code and auditable loops, not adjectives on a slide.
后台任务循环自主跟进每一段关系：判断该不该主动开口、查知识、起草、评审、发送，全程无需人盯。一条入站消息自动触发完整决策闭环。
A background task loop nurtures every relationship on its own — deciding whether to reach out, consulting knowledge, drafting, reviewing and sending, with no human watching. One inbound message triggers the full loop.
自己约束自己：独立评审与认知隔离、事实/压力/知识三道闸门、红线由守门函数与 CI 静态扫描强制。越界的话发不出去，错误也绝不抛给客户。
It governs itself: an independent reviewer with epistemic distance, three gates on fact / pressure / grounding, hard lines enforced by guard functions and a CI lint. Out-of-bounds replies never ship; errors never reach the customer.
单进程托管后台、API、微信回调与多个后台 worker。幂等发件箱自动抢占派发、超时回收重试，长期稳定运转，不重发、不漏发。
A single process hosts the admin, API, WeChat callbacks and multiple background workers. An idempotent outbox claims, dispatches, reclaims and retries on its own — running long-term without double- or missed sends.
演化引擎从历史决策中学习，在影子环境重放重判候选阈值与提示词；只有显著变好、且通过零容忍安全回归门，再经管理员确认才发布。
An evolution engine learns from past decisions, replaying and re-grading threshold and prompt candidates in a shadow. Only a significant gain that clears a zero-tolerance safety gate — and an admin's confirmation — ships.
WeAgent 的「聪明」不是靠堆提示词，而是把心理学与认知科学里被反复验证的方法落进工程：长短期记忆、多轮证据置信、人格画像、意图走势。下面每一项都对应真实代码，也如实说明它到底做了什么、没做什么。
WeAgent's "smarts" don't come from piling up prompts but from engineering well-validated methods from psychology and cognitive science: long/short-term memory, multi-round evidence confidence, personality profiling and intent trajectories. Each maps to real code — and we say plainly what it does and doesn't do.
短期上下文承载眼前对话；长期记忆是一张随时间固化的记忆卡。新观察先进候选池，再由整理 Agent 固化、压缩、判定过时事实——像人脑把经历沉淀成长期记忆。
Short-term context holds the live conversation; long-term memory is a card that consolidates over time. New observations enter a candidate pool, then a consolidator promotes, compacts and deprecates facts — much like the brain settling experience into long-term memory.
对每个客户维度持续累积证据：必须跨多轮反复命中、且由代码侧客观统计到足够「强证据」（锚定客户真实发言），才会被确认。不轻信单次表态，像理性的人一样越看越准。
Evidence for each customer dimension accumulates over time: a signal is only confirmed after repeated hits across rounds plus enough "strong evidence" objectively counted in code (anchored to the customer's own words). It distrusts one-off claims — getting surer the more it sees.
用心理学公认的「大五人格」(OCEAN) 五个维度刻画客户，并按记忆周期留存演化快照。关键的克制：人格只作观测参考、永不直接驱动话术——避免基于标签的偏见，证据不足时置信度直接归零。
It profiles customers along the psychology-standard Big Five (OCEAN) and keeps evolution snapshots per memory cycle. A deliberate restraint: personality is observation-only and never directly drives wording — avoiding label-based bias, with confidence forced to zero when evidence is thin.
把每一轮的推进结果记成时间序列，形成一条意图走势线。这条线真的会改变 AI 的下一步：连续多轮没推进、末轮还转负，就触发换策略或向幕后请示——会复盘、会调整，不一条道走到黑。
Each round's progress outcome is recorded as a time series — an intent trajectory. It genuinely changes the next move: several rounds without progress plus a negative turn triggers a strategy switch or a behind-the-scenes consult. It reflects and adapts rather than pushing blindly.
无论是自动回复还是定时跟进，都不能绕过这条链路。绕过网关，就是 bug。
Auto-reply or scheduled follow-up — nothing bypasses this path. Bypassing the gateway is a bug.
解析消息，定位账号与联系人。仅「已托管」联系人进入决策；依次检查冷却、最小间隔、每日上限、任务过期、作息时段。
Parse, resolve account & contact. Only "managed" contacts proceed; cooldown, min-interval, daily cap, expiry and quiet-hours are checked in order.
目录 → 列切片 → 打开切片 → 取证。LLM 自主工具规划，不用向量库；无依据时只保守回应。
Catalog → list → open slice → cite. The LLM plans tools itself, no vector DB; ungrounded means a conservative reply.
渐进式三档（精简→关系→完整）。Reply Agent 输出回复、运营状态、下一步最优动作与自治模式。
Progressive 3 tiers (lean → relational → full). The Reply Agent emits reply, operation state, next-best-action and autonomy mode.
独立 Review Agent 只看事实面、看不到推理过程，避免自我追认。硬闸拦截、软闸触发一次改写。可选双模并行交叉验证。
A separate reviewer sees only facts, not the reasoning — no self-confirmation. Hard gates block; soft gates trigger one rewrite. Optional dual-model cross-check.
仅「通过」的决策才入出箱，幂等键防重发。发送前二道安全门：再查冷却、用户是否喊停、是否过期。
Only approved decisions enter the outbox, keyed for idempotency. A second safety gate re-checks cooldown, user stop, and staleness before send.
异步分析用户反应（带抢锁），更新意图轨迹、长期记忆与风格指纹，写入决策复盘与运行日志。
Async reaction analysis (lock-guarded) updates intent trajectory, long-term memory and style fingerprint; decision review and run log are persisted.
我们正在用真实大模型，对整个项目逐个业务域做端到端测试：真调用、真对话、真评审、真发送决策。AI 时代的产品，必须用 AI 的方式验证——只看断言是否变绿没有意义，要看大模型在真实业务里到底怎么表现。
We are running end-to-end tests with real LLMs across every business domain of the whole system: real calls, real dialogue, real review, real send decisions. An AI-native product must be verified the AI way — green assertions alone mean little; what matters is how the model actually behaves in real business.
每个域都是一段真实运营剧本，由真实大模型从头跑到尾，发现的问题沉淀为可复现的方法论缺陷，而非对单条对话打补丁。
Each domain is a real operating script run end-to-end by a real LLM; findings are distilled into reproducible methodology gaps, never patched against a single conversation.
维度、人格、状态机、评审闸门、方法论公式全部可配置。系统对行业零假设，未配置时回落到与销售域字节等价的默认画像。
Dimensions, persona, state machine, review gates and methodology formulas are all configurable. Zero assumptions about industry — it falls back to a sales profile byte-for-byte when unconfigured.
客户 / 同行 / 朋友三种关系类型，各自切换主动触达范式：客户全开追单，同行朋友关掉销售漏斗。数字分身因人而异。
Customer / peer / friend each switch the proactive-outreach mode: full follow-up for customers, funnel off for peers and friends. The digital twin adapts per person.
Soul Prompt 表达稳定人格，每个联系人可用自然语言指令微调口吻——更直接、更温和、更技术化。指令置于系统提示末位，优先级最高。
A Soul Prompt holds the stable persona; per-contact natural-language instructions tune the voice — more direct, gentler, more technical. They sit last in the prompt, top priority.
不是话术机器人，而是会记忆、会判断、会克制、会请示的运营官。全程可审计，红线由代码守护。
Not a script bot — an operator that remembers, judges, restrains, and consults. Fully auditable, with hard lines guarded in code.