情报台一号 · 架构图库Desk 1 · Architecture Library
把一团乱麻，剪成一张架构图。Cutting the tangle into one architecture diagram.
新架构、新思想与一线信源按十关入库，原图有版权，公开的一律是结构拆解与要点：七层与十关的完整定义见New architectures, new ideas and frontline sources, registered by the ten gates. Original diagrams are copyrighted; what we publish is the structural breakdown: for the full definitions of the seven layers and ten gates, see the框架规范 →framework specification →
深拆解的读法与首批条目，见 AI 架构研究库 →For how the deep teardowns are written and the first entries, see the Architecture Research Library →
实线为层内七关，虚线为跨层三关（网关、平台、统一与自主）。每格条数按架构卡与新思想合计，跨关条目在每个所属关各计一次；产品拆解 PD-001 在下方专区单列，不计入地图条数。点一关只看那一关的条目。Solid borders are the seven in-layer gates; dashed borders are the three cross-layer gates (Gateway, Platform, Unity and Autonomy). Counts combine architecture cards and new ideas; an entry belonging to several gates is counted once in each of them. Teardown PD-001 sits in its own zone below and is not counted on the map. Click a gate to see only its entries.
从 Chatbot 到企业级系统的全景框架，DeepWing 全部分析的基准。The full map from chatbot to enterprise system; the baseline for all DeepWing analysis.
来源：DeepWingAI · 打开来源Source: DeepWingAI · Open source
关联条目Linked entries暂无None yet
26B 模型拆开摊在 15 台约 55 美元的电视盒子上合跑，约 0.39 token 每秒。A 26B model split across 15 TV boxes at about $55 each, jointly serving at roughly 0.39 tokens per second.
来源：Anthropic Frontier Red TeamSource: Anthropic Frontier Red Team · 打开来源View source
关联条目Linked entriesSP-001
本体在服务器上，迁移安家的是文件包，在 8 台模拟设备间传播。The model stays on a server; a file package migrates and settles across 8 emulated devices.
来源：Anthropic Frontier Red TeamSource: Anthropic Frontier Red Team · 打开来源View source
关联条目Linked entriesSP-001
一个 run 自主拉起最多 1000 个子 Agent，服务端最多 64 个并发，成本随规模乘法增长。One run autonomously spawns up to 1,000 sub-agents, at most 64 concurrent server-side; cost grows multiplicatively with scale.
来源：Anthropic 官方文档（Managed Agents 多智能体编排） · 打开来源Source: Anthropic official documentation (Managed Agents multi-agent orchestration) · Open source
关联条目Linked entriesIDEA-003
切分、向量化、索引、检索、重排、注入上下文的标准链路。The standard chain: chunking, embedding, indexing, retrieval, reranking, context injection.
来源：DeepWingAI · 打开来源Source: DeepWingAI · Open source
关联条目Linked entries暂无None yet
轻量小控制器管派活与喊停，大模型只做推理。对 GPT-4o 的 token 消耗最多降 97%，延迟最多降 93%。Lightweight controllers handle assignment and stopping; big models only reason. GPT-4o token use cut by up to 97%, latency by up to 93%.
来源：arXiv 2610.08155Source: arXiv 2610.08155 · 打开来源View source
关联条目Linked entries暂无None yet
Agent 不直接喊话，往可搜索的共享消息库里写和查，撑到 128 个 Agent 不乱套。Agents coordinate through a searchable shared message store instead of direct chatter; scales to 128 agents.
来源：Meta FAIR / Harvard, arXiv 2026Source: Meta FAIR / Harvard, arXiv 2026 · 打开来源View source
关联条目Linked entries暂无None yet
不要中央编排器，每一步动作拿出来拍卖，Agent 竞价上岗，带预算约束。No central orchestrator: each next action is auctioned, agents bid to take it, with budget constraints.
来源：arXiv 2610.22201Source: arXiv 2610.22201 · 打开来源View source
关联条目Linked entries暂无None yet
多 Agent 系统运行时巡检与当场修正：不等跑完复盘，执行中发现错误即修；论文给出 3,953 组已核修复对（来自 24k 条轨迹）。Runtime inspection and on-the-spot correction for multi-agent systems: errors are fixed as they occur during execution, not in a post-hoc review; the paper reports 3,953 verified repair pairs drawn from 24k trajectories.
来源：arXiv:2610.20780v1（Shen & Dong）Source: arXiv:2610.20780v1 (Shen & Dong) · 打开来源View source
关联条目Linked entriesIDEA-006
用 AP2 签名授权（买什么、向谁买、上限多少）把「身份」和「放行权」分开管：买方在 AWS、卖方在 Cloudflare Workers、授权与验签在 Google Cloud，Stripe 完成测试支付，另配 Trust Index。授权金额即硬上限，与「AI 限薪」线（额度即门禁）和判例库的归因/审计直接呼应。AP2 signed mandates (what may be bought, from whom, and up to how much) manage identity and release rights separately: the buyer runs on AWS, the seller on Cloudflare Workers, authorization and signature verification on Google Cloud, with Stripe completing a test payment, plus a Trust Index. The authorized amount is a hard cap, directly echoing the “AI salary cap” line (budget as access control) and the case library’s attribution and audit concerns.
来源：PRNewswire（GoDaddy、Infoblox、Google、Cloudflare 联合 IMDA 发布） · 打开来源Source: PRNewswire (GoDaddy, Infoblox, Google, Cloudflare with IMDA) · Open source
当约九成代码由 AI 写出后，审查从终点单门改成分层流水线，一次变更依次过七道：提交前自验（代理自查后再由另一代理以全新上下文审 diff）；每个 PR 多代理并行查错，发现经独立投票核实才贴出；安全审查单列，未决问题阻断合并；自动化检查（lint、类型、全平台测试、构建）全绿才放行；人工批准按风险分级（低风险零人、中风险一人、高风险两人，代理的批准不计作人的批准）；合并队列对主干复跑全部检查，新行为先经功能开关小范围放量；快速回滚兜底。另配收敛规则防代理审查打转：三轮上限、后续只审新 diff、按问题类别整类修、已贴问题不重复。官方口径效果：2026 年 6 至 9 月合并 PR 数量翻倍多，回滚占比下降四成多，获实质审查评论的内部 PR 占比由 16% 升至 54%。Once roughly ninety percent of the code is written by AI, review stops being a single gate at the end and becomes a layered pipeline: each change passes seven stages: pre-submit self-checks (the agent verifies itself, then another agent reviews the diff with fresh context); multi-agent review of every PR, with findings posted only after independent vote verification, graded by severity and split into newly introduced vs. pre-existing; a separate security review whose open issues block merging; automated checks (lint, types, full-platform tests, builds) that must all pass; human approval graded by risk (a classifier reads the diff: low risk needs no human, medium one, high two; an agent’s approval never counts as a human’s, and approval binds to one specific commit); a merge queue re-running all checks against the trunk, with new behavior first rolled out behind feature flags; and fast rollback as the backstop. Convergence rules keep agent review from looping: a three-round cap, later rounds review only the new diff, fixes land by problem category, and posted findings are never repeated. Reported effect: merged PRs more than doubled from June to September 2026, the rollback share fell by over forty percent, and the share of internal PRs receiving substantive review comments rose from 16% to 54%.
来源：Anthropic 官方博客（claude.dev）《How the Claude Code team does code review》，2026-10-15Source: Anthropic official blog (claude.dev), “How the Claude Code team does code review”, Oct 15, 2026 · 打开来源View source
关联条目Linked entriesPD-001GOV-005GOV-006GOV-007GOV-008IDEA-009
基座模型冻结不动，只改模型外围的脚手架：系统提示词、护栏、记忆、技能、工具指导、脚本、子智能体、插件八个文件，每次改动都是一个文件 diff。任务内：先跑一遍暴露问题，两个编辑角色各提一版改动，同题重跑做行为对比，再选择性合成；任务交付后：只把有对比支持的教训连同可执行组件存进经验库，供后续任务复用。最要紧的门禁事实：护栏文件本身就在可编辑清单里，针对护栏被改弱的单独复核未见披露。效果数字全部为论文作者自报，无第三方复现。The base model stays frozen; only the scaffold around it is edited, as eight files: system prompt, guardrails, memory, skills, tool guidance, scripts, sub-agents and plugins, so every change is a file diff. Within a task, a reference run exposes problems, two editor roles each propose a change, the task is rerun to compare behavior, and the changes are composed selectively. After delivery, only lessons backed by those comparisons are kept, with their executable components, in an experience library for later tasks. The key gate fact: the guardrails file is itself on the editable list, and no separate review of a weakened guardrail is disclosed. All performance figures are self-reported by the paper’s authors, with no third-party replication.
条目 v1.0（2026-10-27 核查）：一手已核，论文与代码仓双来源直读；承重判断 51 条，其中未知 10 项（19.6%），包括护栏改弱无单独复核、经验库无淘汰机制、生产运行监测未披露。论文自陈的失效方式：最终选择选到更弱的候选；错误前提在候选间传播；不可靠教训入库后跨任务放大；改进回路把自身约束改弱；递归预算越大成本越高且无最优停点。适用前提：问题以程序性为主、任务可同题重跑、执行端网络与沙箱另行收紧；代码仓声明不是 Google 官方支持的产品。Entry v1.0 (checked 2026-10-27): primary-source verified against both the paper and the code repository; 51 load-bearing claims, 10 of them unknown (19.6%), including no separate review of weakened guardrails, no retirement mechanism for the experience library, and no disclosed production monitoring. Failure modes the paper itself reports: the final selector picks a weaker candidate; a wrong premise spreads across candidates; an unreliable lesson enters the library and is amplified across tasks; the improvement loop weakens its own constraints; and cost rises with the recursion budget with no optimal stopping point. It fits when problems are mainly procedural, tasks can be rerun, and the executor’s network and sandbox are locked down separately; the repository states it is not an officially supported Google product.
来源：论文 arXiv:2610.14224《Harness^2: Recursive Agent Harnessing for an Open World》，2026-10-09 · 打开论文；代码仓 google-research/harness2（Apache-2.0） · 打开代码仓Sources: paper arXiv:2610.14224, “Harness^2: Recursive Agent Harnessing for an Open World”, Oct 9, 2026 · Open paper; code repository google-research/harness2 (Apache-2.0) · Open repository
Google 把智能体治理拆成四问：它是谁、能做什么、做了什么、绝不能碰什么，每一问都有成文的控制。每个智能体有密码学证明的独立身份（SPIFFE 标准）；全部进出与智能体之间的流量都过 Agent Gateway，策略写一次全员适用；任务在每会话一台的单租户 microVM 沙箱里跑，沙箱内不放任何长效凭证，出站时由网关验证用户委派、智能体身份与节点身份三项后，才在链路上换入后端凭证；每个动作写入审计并归因到智能体；项目级支出上限一触发，该项目的智能体即暂停。最要紧的门禁事实：凭证永不进沙箱，提示注入的最坏后果从「凭证被盗、横向移动」降为「在自身权限内做错事」。智能体本体尚未宣布正式可用，多项治理扩展仍是预览，这里记的是成文设计，不是验收结果。Google breaks agent governance into four questions: who the agent is, what it may do, what it did, and what it must never touch, and each question has a documented control. Every agent gets its own cryptographically attested identity (on the SPIFFE standard); all traffic in, out and between agents passes through Agent Gateway, where a policy written once applies to every agent; tasks run in a single-tenant microVM sandbox, one per session, that holds no long-lived credentials, and on the way out the gateway checks three identities (the user's delegation, the agent and the node) before it swaps the backend credential into the request; every action is written to an audit trail attributed to the agent; and a project-level spend cap pauses that project's agents when it trips. The key gate fact: credentials never enter the sandbox, so the worst case of a prompt injection drops from stolen credentials and lateral movement to the agent doing the wrong thing within its own permissions. The agent itself has not been declared generally available and several governance extensions are still in preview, so this records the documented design, not verified delivery.
条目 v1.3（2026-10-28 核查，2026-10-29 修订）：一手已核，三篇 Google Cloud 官方博客全文直读，另以官方文档与定价页补核；承重判断 62 条，其中未知 12 项（19.4%），包括 Smart Routing 按什么规则分流、子智能体的步数与时长上限、四类记忆的留存与删除规则、网关策略的临时豁免与误配回滚、审计记录能否防篡改。厂商自报的数字（token 降价、TPU 性价比、基准与客户效果）一律单独标为厂商声明，不作结论依据。失效方式：同事型智能体原样继承组织里过度分享的文件；邮件与文档里藏的指令劫持智能体；执行环境隔离被突破；借体外服务绕过出站限制；长任务支出失控，或上限触发时整个项目停摆；共享卷成为控制面之外的协调通道。适用前提：组织已在 Google 生态内运转、能接受预览期产品、部署前先清理分享权限、成本能按项目落账。Entry v1.3 (checked 2026-10-28, revised 2026-10-29): primary-source verified against the full text of three official Google Cloud blog posts, with a further check against official documentation and the pricing page; 62 load-bearing claims, 12 of them unknown (19.4%), including how Smart Routing decides, step and duration limits for sub-agents, retention and deletion rules for the four kinds of memory, exemptions and rollback for gateway policies, and whether audit records are tamper-proof. Vendor-reported figures (token price cuts, TPU price-performance, benchmarks and customer results) are all labelled as vendor claims and are not used as grounds for any conclusion. Failure modes: coworker agents inherit whatever the organization has over-shared; instructions hidden in email and documents hijack the agent; execution isolation is breached; outside services are used to get around egress limits; long tasks run up costs, or a tripped cap stalls a whole project; and shared volumes become a coordination channel outside the control plane. It fits organizations already running on Google, able to accept a preview-stage product, willing to clean up sharing permissions before deployment, and able to book costs by project.
来源：Thomas Kurian《Gemini at Work 2026: Introducing Gemini agent》，2026-10-08 · 打开发布文；《Introducing Agent Compute Engine and expanding agent governance》，2026-10-22 · 打开治理扩展文；《Agent Compute Engine deep dive》，2026-10-22 · 打开深挖文；另据 Google Cloud 官方文档与定价页（2026-10-28 补核）Sources: Thomas Kurian, “Gemini at Work 2026: Introducing Gemini agent”, Oct 8, 2026 · Open launch post; “Introducing Agent Compute Engine and expanding agent governance”, Oct 22, 2026 · Open governance post; “Agent Compute Engine deep dive”, Oct 22, 2026 · Open deep dive; plus Google Cloud official documentation and pricing page (checked Oct 28, 2026)
关联条目Linked entriesINC-014INC-015INC-033INC-038AGT-001AGT-002IDEA-001IDEA-003IDEA-005IDEA-007
当前筛选下这一区没有条目，换一关或换个类型看看。Nothing in this zone under the current filter. Try another gate or type.
按层拆开看：门禁前移的提交前自验（代理自查，加独立代理以全新上下文复审 diff）；盯梢层的 PR 级 AI 审查（多代理并行查错，独立投票核实才贴出，按重要、小瑕疵、既有三档标注）；门禁层的硬阻断（安全审查未决问题不许合并；lint、类型、全平台测试、构建全绿才放行）；门禁层治理（人工批准按风险分级，分类器读变更定级，低风险零人、中风险一人、高风险两人，代理的批准不计作人的，批准只对特定提交有效）；平台层（合并队列对主干复跑全部检查，新行为经功能开关先内部放量）；盯梢层兜底（快速检测与回滚、值班轮换、事故复盘反哺审查者评测集）。对应架构条目 ARC-011，对应想法条目 IDEA-009。Taken layer by layer: pre-submit self-checks push the gate forward (the agent checks itself, then an independent agent re-reviews the diff with fresh context); PR-level AI review sits in Observability (parallel agents hunt for bugs, findings are posted only after independent vote verification, labeled important, nit, or pre-existing across three grades); hard blocks sit in Access Control (unresolved security findings forbid merging; lint, types, full-platform tests, and builds must all be green); governance also sits in Access Control (human approval graded by risk: a classifier reads the change: low risk needs no human, medium one, high two; an agent’s approval never counts as a human’s, and approval binds to one specific commit); the Platform layer re-runs everything against the trunk in the merge queue and releases new behavior behind feature flags internally first; and Observability closes the loop (fast detection and rollback, on-call rotation, incident reviews feeding the reviewer eval set). Linked architecture entry ARC-011; linked idea entry IDEA-009.
关键事实：约九成 Claude Code 代码由 Claude 自己写出；2026 年 6 至 9 月合并 PR 数量翻倍多，回滚占比下降四成多；单日变更量约由 2,500（5 月）升至 30,000 以上（9 月）；获实质审查评论的内部 PR 占比由 16% 升至 54%；曾有单个 PR 经历五十多轮审查，git blame 显示约 59% 被反复标记的问题在第一轮版本中已存在；收敛规则为三轮上限、后续只审新 diff、按问题类别整类修、已贴出的问题不重复；出过事的变更做成评测集，审查者每日回放打分；按回滚率与事故率自动收紧批准的熔断器尚不存在，是官方自陈的待建项。Key facts: roughly ninety percent of Claude Code’s code is written by Claude itself; merged PRs more than doubled from June to September 2026 while the rollback share fell by over forty percent; daily changes rose from about 2,500 (May) to over 30,000 (September); the share of internal PRs receiving substantive review comments rose from 16% to 54%; one PR went through fifty-plus review rounds, and git blame showed about 59% of its repeatedly flagged issues were present in the first version; the convergence rules are a three-round cap, later rounds review only the new diff, fixes land by problem category, and posted findings are never repeated; changes that caused incidents become an eval set the reviewers replay and score daily; a circuit breaker that automatically tightens approvals by rollback and incident rates does not exist yet, an acknowledged to-build item in the official account.
来源：Anthropic 官方博客（claude.dev）《How the Claude Code team does code review》，2026-10-15Source: Anthropic official blog (claude.dev), “How the Claude Code team does code review”, Oct 15, 2026 · 打开来源View source
关联条目Linked entriesARC-011IDEA-009GOV-005GOV-006GOV-007GOV-008
关注点：知识的版本与引用溯源、权限继承、检索与注入的日志留痕。Focus: knowledge versioning and citation traceability, permission inheritance, and logged retrieval and injection.
关注点：工具动作的边界、被拒即停、人工接管与快速回滚。Focus: boundaries on tool actions, stop when refused, human takeover and fast rollback.
关注点：网关代持凭证、编排托管、用量与成本计量。Focus: gateway-held credentials, hosted orchestration, and usage and cost metering.
关注点：身份与授权分离、记忆的可见与可删、跨应用动作审计。Focus: identity separated from authorization, visible and deletable memory, and cross-app action audit.
收录标准：只拆有官方技术披露的产品，每篇用同一个六段模板。Inclusion rule: we only tear down products with official technical disclosures, and every teardown uses the same six-part template.
出站流量带身份标记与出处，配限流，事故才可归因。Outbound agent traffic carries identity and provenance markers with rate limits, so incidents can be attributed.
来源：Wikimedia Foundation 官方披露（Diff 博客，2026-10-05） · 打开来源Source: Wikimedia Foundation official disclosure (Diff blog, 2026-10-05) · Open source
关联条目Linked entries暂无None yet
AI 生成文本加机器可读标记，出处治理变成基础设施。Machine-readable marks on AI-generated text turn provenance into infrastructure.
来源：OpenAI 官方博客（欧盟文本出处规则，textGrain） · 打开来源Source: OpenAI blog (EU text provenance rules, textGrain) · Open source
关联条目Linked entries暂无None yet
编排层正被平台收走，主 Agent 在服务端拉起子 Agent 群。Platforms are absorbing the orchestration layer, with lead agents spawning sub-agent fleets server-side.
来源：OpenAI 与 Anthropic 官方文档 · 打开来源（OpenAI） · 打开来源（Anthropic）Source: OpenAI and Anthropic official documentation · Open source (OpenAI) · Open source (Anthropic)
关联条目Linked entriesARC-004
授权被拒后熔断并上报，不许换路重试。DeepWing 从判例中提炼。After a denial, halt and report instead of rerouting. Distilled by DeepWing from the case library.
真密钥交网关托管，Agent 只持限额的虚拟密钥。DeepWing 七层框架，网关层。Real credentials stay in the gateway; agents hold budgeted virtual keys. From DeepWing’s seven layers, the gateway layer.
来源：深翼 DeepWing 自有框架（七层框架 · 网关关） · 打开来源Source: DeepWing framework (seven layers · Gateway gate) · Open source
关联条目Linked entries暂无None yet
错误处理从「事后复盘」前移到「运行时纠偏」；接 CCA 多智能体线与判例库（运行时修复 vs 事后验尸）。来自 ARIA 论文（arXiv:2610.20780v1）。Error handling moves from post-hoc review to correction during execution; it ties into the CCA multi-agent line and the case library (runtime repair vs. post-mortem). From the ARIA paper (arXiv:2610.20780v1).
来源：ARIA 论文（arXiv:2610.20780v1） · 打开来源Source: the ARIA paper (arXiv:2610.20780v1) · Open source
关联条目Linked entriesARC-009
Agent 的身份与放行权分开管，授权用可签名、可限额的 mandate 表示，授权金额即硬上限；与「AI 限薪」线（额度即门禁）和判例库的归因/审计直接呼应。跨厂商试点中（GoDaddy、Infoblox、Google、Cloudflare 联合 IMDA，GovWare 2026）。An agent’s identity and its release rights are managed separately, with authorization expressed as a signed, capped mandate: the authorized amount is a hard cap. It echoes the “AI salary cap” line (budget as access control) and the case library’s attribution and audit concerns. In cross-vendor pilot (GoDaddy, Infoblox, Google, and Cloudflare with IMDA, GovWare 2026).
记忆变成可计量的成本项。AWS AgentCore Memory 按写入、读取、存储分开计价，记忆不再只是架构概念，而是可计量、可预算的成本项；对应成本记录 COST-004。Memory becomes a measurable cost line. AWS AgentCore Memory prices writes, reads, and storage separately: memory is no longer just an architecture concept, but a meterable, budgetable cost item; see Cost Index COST-004.
来源：AWS AgentCore Memory 定价，对应成本记录 COST-004 · 打开来源Source: AWS AgentCore Memory pricing; see cost index COST-004 · Open source
关联条目Linked entriesCOST-004
把历史上引发事故的 PR 连同其回滚与修复收成评测集，由熟悉代码的工程师逐条确认缺陷是什么、在哪一层；审查系统（提示词、工具、模型）每次改动先在这套题上打分，评测每日跑，新事故持续扩充题库。事故复盘固定追问哪一层本该拦住：与本站判例库同构，是盯梢层/门禁层的反馈闭环。来自 Anthropic Claude Code 团队代码审查实践（官方博客，2026-10-15）。Historical PRs that caused incidents, together with their rollbacks and fixes, become the reviewers’ exam questions; engineers who know the code confirm, item by item, what the defect was and at which layer; every change to the review system (prompts, tools, model) is scored on this set first, the eval runs daily, and new incidents keep expanding the question bank. Every incident review asks which layer should have caught it: the same shape as this site’s case library, a feedback loop for the Observability and Access Control layers. From the Anthropic Claude Code team’s code review practice (official blog, Oct 15, 2026).
来源：Anthropic 官方博客（2026-10-15） · 打开来源Source: Anthropic official blog (2026-10-15) · Open source
当前筛选下这一区没有条目，换一关或换个类型看看。Nothing in this zone under the current filter. Try another gate or type.
七层是架构模型，一个系统由哪七层组成；十关是评估检查点，拿这张地图去量一个系统时，要逐一过堂的十个关口。十关等于七层，再加上请求进门时的网关、托底的平台，以及跨层的统一与自主。The seven layers are the architecture model: what a system is made of. The ten gates are the evaluation checkpoints a system is walked through, one by one. The ten gates are the seven layers, plus the gateway at the request entrance, the platform underneath, and unity and autonomy across layers.
| 十关Gate | 对应七层Maps to | 类型Kind |
|---|---|---|
| 网关Gateway | 七层之外，请求入口Outside the seven layers: the request entrance | 跨层Cross-layer |
| 门禁Access Control | 门禁层Access Control layer | 层内In-layer |
| 编排Orchestration | 编排层Orchestration layer | 层内In-layer |
| 知识Knowledge | 知识层Knowledge layer | 层内In-layer |
| 记忆Memory | 记忆层Memory layer | 层内In-layer |
| 工具Tool Use | 工具层Tool layer | 层内In-layer |
| 模型Model | 模型层Model layer | 层内In-layer |
| 盯梢Observability | 盯梢层Observability layer | 层内In-layer |
| 平台Platform | 七层之外，运行底座Outside the seven layers: the runtime base | 跨层Cross-layer |
| 统一与自主Unity and Autonomy | 七层之外，跨层治理Outside the seven layers: cross-layer governance | 跨层Cross-layer |
