组织级第二大脑：构建从专家学习的 AI

• 我们构建了一个 AI 代理，作为特定领域的辅助专家，让组织中的任何人都能轻松获取、共享并在此基础上构建深厚的专业知识，同时将这些知识保存下来。
  We’ve built an AI agent that acts as a secondary expert for a given domain, making deep specialist knowledge readily available and preserved for anyone in an organization to access, share, and build upon.
• This is not a typical domain-specific agent. Its novelty comes from integrating two layers: 
  • A structured, auditable knowledge architecture separates what the agent knows from how it reasons. 
  • A self-improvement loop then compiles expert feedback into verified, regression-tested updates without model retraining. 
• 这两个层共同将一次性的专家修正转化为永久积累、不断复利的机构记忆，而这一模式旨在推广到由可检索文本而非模型权重所支配的其他领域。
  Together, these two layers turn one-off expert corrections into permanent, compounding institutional memory, and the pattern is designed to generalize to other domains governed by retrievable text rather than model weights.
• 该系统为 Meta 的领域主题专家（SME）节省了大量时间，使他们能够更加专注于那些最需要运用其专业知识的工作。
  This system is saving domain subject matter experts (SME)s at Meta substantial time, allowing them to focus more on the work where their knowledge matters most.

许多大型组织在专业知识方面都面临同样的问题。虽然其中一部分以模型、操作手册、检查清单和框架的形式被记录下来，但最有价值的专业知识存在于人们的头脑中，很少被持久地记录在任何地方。例如，在合规领域，同类型的问题可能会出现在数百次产品审查中，专家评估需要数天的人工研究，而评估之间的不一致会给组织带来实际风险。Many large organizations have the same problem when it comes to specialist knowledge. While some of it is written down in the form of models, playbooks, checklists, and frameworks, the most valuable specialist knowledge lives in people’s heads and rarely gets captured anywhere durable. In compliance domains, for example, the same types of questions can arise across hundreds of product reviews, expert assessments take days of manual research, and inconsistency between assessments creates real organizational risk.

专家花在回答常规问题上的时间多于处理真正新颖且模糊、最需要其判断力的工作，这种情况并不少见。我们需要构建能够捕捉组织专家推理方式的系统，并将这些知识提供给所有需要的人，从而让专业知识更易于分享、借鉴和保存。为了解决这一挑战，我们将组织智慧编码为一个面向特定合规领域的 AI 智能体。该智能体结合了一个充当组织“第二大脑”的知识系统、一个模拟领域专家实际思维方式的推理层，以及一个能够持续累积专家投入并永久提升系统能力的自动化改进流水线。这些模式可以推广到任何拥有深厚专业知识的企业领域，无论是金融、安全还是工程。It’s not uncommon for experts to spend more time answering routine questions than on genuinely novel and ambiguous work where their judgment matters most. We need systems that can capture how an organization’s experts reason and make that knowledge available to everyone who needs it, so that expertise is easier to share, build on, and preserve.

We set about solving this challenge by codifying institutional intelligence into an AI agent for a specific compliance domain. The agent combines a knowledge system that acts as the organization’s “second brain,” a reasoning layer that mirrors how domain experts actually think, and an automated improvement pipeline that compounds expert effort permanently. The patterns generalize to any enterprise domain with deep specialist knowledge, whether that is finance, security, or engineering. 

架构概览The Architecture at a Glance

现成的 LLM 提供了坚实的基础，但要在专业领域充分发挥作用，通常还需要更深入的机构背景知识。缺乏这种基础时，通用模型的价值会受到限制，因为它无法区分组织能够做什么（对一般信息的总结）与组织应当考虑做什么（基于历史立场、公司方向、业务背景等）。在高风险领域，弥合这一差距需要向模型提供组织自身的知识和优先事项，使其分析能够反映组织实际进行推理的方式。Off-the-shelf LLMs provide a strong foundation, but they often need deeper institutional context to be fully effective in specialist domains. Without that grounding, a general purpose model has limited value given it will not be able to distinguish between what an organization could do (a summary of general information) and what it should consider doing (based on historic positions, company direction, business context, etc.). In high-stakes domains, closing this gap requires supplying the model with the organization’s own knowledge and priorities so its analysis reflects how the organization actually reasons.

我们设计的系统由四个层组成，每一层都解决一个独特的问题：The system we’ve designed has four layers, each solving a distinct problem:

这些层彼此依赖。知识系统的文件结构使自动化编辑成为可能。推理层的明确流程使失败归因变得可行。评估框架会对每项变更进行把关。而改进循环则将反馈同时传回知识和推理。移除任何一层，其他层都会退化。These layers depend on each other. The knowledge system’s file structure makes automated editing possible. The reasoning layer’s explicit procedures make failure attribution tractable. The evaluation framework gates every change. And the improvement loop feeds back into both knowledge and reasoning. Remove any one layer and the others degrade.

大型组织可以在专家工作的过程中积累数千份文档。人们很容易将这些文档视为组织知识，但真正的知识是隐性的：专家如何进行推理、优先考虑什么，以及如何解决歧义。在推理时检索文档片段的智能体，必须在每次运行时都从原始资料中重新推导出这种推理过程，这既缓慢、容易出错，又缺乏一致性。Large organizations can accumulate thousands of documents as a byproduct of expert work. It is tempting to treat those documents as organizational knowledge, but the real knowledge is implicit: how experts reason, what they prioritize, and how they resolve ambiguity. An agent that retrieves document chunks at inference time has to re-derive that reasoning from raw sources on every run, which is slow, error-prone, and inconsistent.

我们会提前将这些隐性知识显性化。一个长期运行的离线流程会对源文档进行推理，并将其提炼为结构化知识文件——其中包含组织如何理解其领域的精选陈述，并将约束条件、边界以及路由影响转化为机器可读的形式。We make that implicit knowledge explicit ahead of time. A long-running offline process reasons through source documents and distills them into structured knowledge files – curated statements of how the organization interprets its domain, with constraints, boundaries, and routing implications made machine-readable. 

最重要的是，这些知识随后构成了一个反馈循环的基础，使智能体能够从人类专家的反馈中学习并加以实施，而无需重新训练底层模型。Most significantly, that knowledge then forms the basis of a feedback loop that allows the agent to learn from and implement feedback from human experts without the underlying model having to be retrained.

业界已经逐渐形成了类似的理念。Andrej Karpathy 的 LLM Wiki 将智能体知识组织成一个可导航的文件图谱，而 Google 的 Open Knowledge Format 则对其进行标准化，以实现跨智能体互操作性。共同的认识是，知识应当预先提取、明确结构化，并逐步披露，而不是在每次查询时重新推导。我们将这些原则进一步扩展为一个对引用忠实度和机构一致性有着不可妥协要求的系统，并按照严格的分类体系组织了 200 多个文件：The industry has converged on a similar idea. Andrej Karpathy’s LLM Wiki structures agent knowledge as a navigable graph of files, and Google’s Open Knowledge Format standardizes this for cross-agent interoperability. The shared insight is that knowledge should be pre-extracted, explicitly structured, and progressively disclosed rather than re-derived on every query. We extended these principles into a system where citation fidelity and institutional consistency are non-negotiable, organizing 200+ files into a strict taxonomy:

• 立场文件记录权威性的组织立场：组织决定如何解读某一特定领域问题，以及其中的约束、边界条件和可由机器执行的路由影响，这些内容会告知推理层何时应用该立场。
  Position files capture authoritative organizational stances: how the organization has decided to interpret a given domain question, along with its constraints, boundary conditions, and machine-actionable routing implications that tell the reasoning layer when to apply it.
• 分类体系和词汇文件充当权威术语表，用于记录组织描述其领域时所使用的术语，例如实体类型、活动类别和分类层级。每个文件都作为单一事实来源进行维护，以确保代理与组织使用一致的语言。
  Taxonomy and vocabulary files act as an authoritative glossary for the terms the organization uses to describe its domain, such as entity types, activity categories, and classification tiers. Each is maintained as a single source of truth so the agent and the organization use language consistently.
• 路由索引将输入特征映射到相关位置和程序，在不单独依赖嵌入相似度的情况下确定哪些文件适用。这使检索过程具有确定性且可审计。
  Routing indexes map input characteristics to the relevant positions and procedures, determining which files apply without relying on embedding similarity alone. This makes retrieval deterministic and auditable.
• 网关文件定义了代理在进入某个分析领域前必须通过的门槛测试，以防止其在不适用的地方运用专业知识。
  Gateway files define threshold tests the agent must pass before entering an analytical domain, preventing it from applying specialized knowledge where it does not belong.

每个文件都在 YAML frontmatter 中声明其依赖项（depends_on）和使用者（referenced_by），从而形成一个双向依赖图。当某个文件发生变化时，你可以准确追踪其他可能受影响的内容；在自我改进循环提出自动化编辑时，这一点至关重要。Every file declares its dependencies (depends_on) and consumers (referenced_by) in YAML frontmatter, forming a bidirectional dependency graph. When one file changes, you can trace exactly what else might be affected, which matters when the self-improvement loop proposes automated edits.

按密度和使用频率组织知识Organizing Knowledge by Density and Usage Frequency

一个关键的架构决策是如何在经过整理的 Wiki 与补充检索（RAG）之间划分知识。我们根据信息密度和预期使用频率进行划分。A key architectural decision is how to partition knowledge between the curated wiki and supplementary retrieval (RAG). We split on information density and expected usage frequency.

高密度、经常被引用的资料应放入维基：这些经过提炼的文件记录了组织如何进行思考，例如立场、决策框架、边界示例和战略解读。智能体几乎每次交互都会查阅这些资料。由于它们承载着组织不断演变的思维方式，因此需要保持最新，而维基的结构使其易于更新、版本管理和验证。High-density, frequently referenced sources go into the wiki: Distilled files capturing how the organization reasons, such as positions, decision frameworks, boundary examples, and strategic interpretations. The agent consults these on nearly every turn. Because they encode the organization’s evolving thinking, they need to stay current, and the wiki structure makes them easy to update, version, and validate.

通过语义或词法搜索（RAG）提供稀疏且与具体情境相关的资料：这些文档在适用时非常重要，但在大多数运行中并不需要详细查阅，例如详细的参考资料、单个产品的规格说明、历史决策记录以及小众的外部知识。将它们全部加载到 wiki 中会使系统臃肿，并分散注意力。Sparse, situationally relevant sources are served through semantic or lexical search (RAG): documents that matter deeply when they apply but are not needed in detail on most runs, such as detailed reference material, individual product specifications, historical decision records, and niche external knowledge. Loading all of them into the wiki would bloat the system and dilute attention.

结果是，智能体的核心推理始终建立在最精细、最新的组织知识之上，同时在具体情境需要时，仍然能够获取辅助证据。这种结合打造了一个组织的“第二大脑”，它编码的是组织如何解读和应用信息，而不仅仅是信息的存放位置。The result is that the agent’s core reasoning is always grounded in the most refined, current organizational knowledge, while it can still reach for supporting evidence when a scenario demands it. The combination produces an organizational second brain that encodes how the organization interprets and applies information, rather than only where to find it.

通过可组合配方实现专家级推理Expert Reasoning via Composable Recipes

仅有知识是不够的。领域专家并不只是回忆事实，他们遵循结构化的方法论：金融分析师会逐步运用估值模型，安全工程师则会遵循威胁建模流程。挑战在于将这些方法论以一种 LLM 能够可靠执行的形式记录下来。Knowledge alone is not enough. Domain experts do not simply recall facts, they follow structured methodologies: a financial analyst works through a valuation model step by step, a security engineer follows a threat modeling procedure. The challenge is capturing those methodologies in a form an LLM can execute reliably.

我们通过称为 recipe 的可组合过程来解决这一问题。知识文件是声明式的，而 recipe 则是命令式的。每个 recipe 都规定了一个多步骤的分析工作流，明确首先检查什么、在每个步骤加载哪些知识、遵循哪些决策流程，以及什么才算完成分析。We solve this with composable procedures we call recipes. Where knowledge files are declarative, recipes are imperative. Each one prescribes a multi-step analytical workflow, specifying what to examine first, which knowledge to load at each step, what decision procedures to follow, and what constitutes a complete analysis.

关键的设计选择是将智能体所知内容与其推理方式分离。配方引用知识文件，但不包含任何领域事实；知识文件阐述立场，但不规定任何操作流程。这意味着：The critical design choice is separating what the agent knows from how it reasons. Recipes reference knowledge files but contain no domain facts; knowledge files state positions but prescribe no procedures. This means:

• 添加一个组织职位意味着添加一个知识文件并更新路由索引。无需更改配方。
  Adding an organizational position means adding a knowledge file and updating a routing index. No recipe changes.
• 修复代理方法论中的缺陷意味着编辑一份配方。知识文件不会发生变化。
  Fixing a flaw in the agent’s methodology means editing a recipe. No knowledge files change.
• 失败可以明确归因于某一层。是知识有误，还是流程有误？
  Failures attribute cleanly to one layer. Was the knowledge wrong, or the procedure?

配方可以组合成流水线，就像主厨为一场晚餐服务准备的主配方会将各个组成部分（酱汁、主菜蛋白、装饰）交由相应的子配方处理，而不会在自身包含这些细节。我们的顶层路由配方会检查输入内容，并选择要调用哪些下游配方，每个下游配方负责一个分析阶段。Recipes compose into pipelines, much like a head chef’s master recipe for a dinner service delegates to sub-recipes for each component (the sauce, the protein, the garnish) without containing those details itself. Our top-level routing recipe examines the input and selects which downstream recipes to invoke, each handling one analytical phase.

这同样是实现渐进式披露的关键。与其预先加载一套涵盖所有可能场景的庞大指令集，不如让配方中的每个步骤仅携带与该阶段相关的指令和知识。早期版本使用单一的扁平指令文件，并通过语义搜索加载所有来源，导致每次运行都将大量相关性不一的文件提取到上下文窗口中。重构为由配方驱动的阶段后，每次查询只会触及一小部分有针对性的内容，使每轮消耗的 token 减少了约 80%。上下文窗口是有限的，注意力也会随着内容量增加而下降，因此在恰当的时间提供恰当的指令，能够直接提升推理质量。This is also what enables progressive disclosure. Rather than front-loading a monolithic instruction set covering every possible scenario, each recipe step carries only the instructions and knowledge relevant to that phase. Early versions used a single flat instruction file and loaded all sources via semantic search, pulling a large volume of mixed-relevance files into the context window on every run. After restructuring into recipe-driven stages, each query touches only a small, targeted subset, cutting tokens consumed per turn by around 80%. Context windows are finite and attention degrades with volume, so delivering the right instructions at the right time directly improves reasoning quality.

让人类掌控局面Keeping Humans in Control

人类专家始终掌控着这一系统。该代理能够加速并结构化他们的工作，但不会取代他们的判断，也不会取代他们对结果的决策权。Human experts stay in control of this system throughout. The agent accelerates and structures their work; it does not replace their judgment or their authority over the outcome. 

我们通过两种机制来强制执行这一点：We enforce this through two mechanisms: 

检查点是在分析过程中定义的节点，智能体会在继续进行之前呈现其中间推理，供专家审查；专家可以确认、纠正或引导其方向。Checkpoints are defined points in the analysis where the agent surfaces its intermediate reasoning for expert review before proceeding, and the expert can confirm, correct, or redirect. 

当代理遇到真正的歧义时就会触发升级，无论这种歧义源于输入信息不充分，还是证据支持不止一种合理解读。它不会强行得出结论，而是将问题交给专家，由专家的选择决定分析所采取的路径。Escalations trigger when the agent hits genuine ambiguity, whether from underspecified inputs or evidence that supports more than one defensible reading. Rather than forcing a resolution, it hands the question to the expert, whose choice determines the path the analysis takes.

检查点和升级机制同时发挥三种作用：Checkpoints and escalations serve three purposes simultaneously:

1. 质量与方向控制：专家会在错误向下游累积之前及时发现，并让分析沿着他们认为最相关的路径进行。
   Quality and direction control: Experts catch errors before they compound downstream, and keep the analysis on the path they consider most relevant.
2. 训练信号：每一次纠正和每一次升级都会成为自我改进循环的输入。
   Training signal: Every correction and every escalation becomes input for the self-improvement loop.
3. 信任校准：专家通过观察智能体的推理过程，而不仅仅是最终输出，并通过看到它主动标记不确定性而非加以掩饰，逐步建立信心。
   Trust calibration: Experts build confidence incrementally by observing the agent’s reasoning rather than just its final output, and by seeing it flag uncertainty instead of masking it.

决策越重要，这一点就越关键；因此，我们建议在合规、金融风险评估、安全审查和工程安全等领域，默认始终让人工参与其中。The more consequential the decision, the more this matters, which is why we recommend keeping a human in the loop by default across domains like compliance, financial risk assessment, security review, and engineering safety.

自我改进飞轮The Self-Improvement Flywheel

我们认为，自我改进飞轮是该系统最具特色的部分。结构化知识系统和可组合的配方使系统对人类和智能体都清晰易懂、可测试且模块化，但相互依赖的文件数量之多，使得手动维护无法扩展。当领域专家向智能体提供反馈时，这些反馈必须被转化为对文件的精确修改。这个过程可能需要数周时间，因为它要求理解完整的依赖关系图、确认没有其他部分遭到破坏，并验证修复确实有效。We consider the self-improvement flywheel to be the most distinctive part of this system. While the structured knowledge system and composable recipes produce a system that is legible by humans and agents, testable, and modular, the number of interdependent files make manual maintenance impossible to scale. When domain experts provide feedback to an agent that feedback has to be translated into precise file edits. This process can take weeks because it requires understanding the full dependency graph, verifying nothing else breaks, and validating the fix actually works.

已有大量工作致力于解决智能体如何存储、检索和更新知识的问题，涵盖从 RAG 记忆系统到模型权重知识编辑等方向。但对于如何在基于文档的机构知识库不断扩展、专家观点不断演变的过程中保持其正确性，所投入的关注却少得多。能够自动起草自身修正方案的智能体正变得越来越常见；但据我们所知，还没有人将这种程度的验证严谨性应用于无需重新训练模型的结构化知识库。There has been a large body of work – from RAG memory systems to model-weight knowledge editing – devoted to addressing how agents store, retrieve, and update knowledge. But much less attention has gone toward keeping a document-based institutional knowledge base correct as it grows and as expert positions evolve. Agents that auto-draft their own fixes are increasingly common; but we haven’t seen this level of validation rigor applied to a structured knowledge base without model retraining. 

我们将这项维护视为一个编译问题，并对其进行自动化处理。每次专家修正都要经过四个阶段：We treat that maintenance as a compilation problem and automate it. Every expert correction moves through four phases:

1. 将专家反馈诊断为可执行的问题，并找出其根本原因。
   Diagnose expert feedback into actionable issues with their root cause.
2. 将问题汇编成经过最小化验证的编辑。
   Compile issues into minimal verified edits.
3. 验证修复有效且不会引入回归。
   Validate that fixes work without regressions.
4. 请领域专家对其进行审查。
   Have domain experts review them.

循环完成后，将刚刚修复的问题添加到回归测试套件中，以确保未来的更新保留此行为。Once the loop completes, the regression test suite is enriched with the issue that was just fixed so that future updates preserve this behavior.

诊断：将每项修正归因于根本原因Diagnosis: Attributing Each Correction to a Root Cause

原始专家反馈来自领域主题专家与代理交互并提供修正意见的对话记录。诊断阶段会从这些对话中提取结构化信号。Raw expert feedback comes from conversation traces where domain SMEs interacted with the agent and provided corrections. The diagnosis phase extracts structured signals from these conversations.

我们的第一种方法根据对话形式对反馈进行分类。如果专家提供了信息，那一定是知识缺口；如果专家将代理引向其他方向，那一定是流程问题。事实证明，这种启发式方法并不奏效，因为对话形式并不能很好地代表根本原因。专家纠正某个结论时，可能是在揭示知识缺口、操作步骤的缺陷，或是真实存在的歧义。Our first approach classified feedback by conversational form. If the expert provided information, it must be a knowledge gap; if they redirected the agent, it must be a procedure problem. This heuristic failed because conversational form is a poor proxy for root cause. An expert correcting a conclusion might be exposing a knowledge gap, a recipe flaw, or a genuine ambiguity.

这种工作方法将提取与分类分开。首先，从专家那里提取每一条实质性信号，以及代理的完整知识清单（加载的每个文件、加载时间及其使用方式）。其次，阅读实际的知识文件，并应用单一的归因测试：代理是否可能仅凭其源材料得出正确结论？The working approach separates extraction from classification. First, extract every substantive signal from the expert alongside the agent’s full knowledge manifest (every file loaded, when, and how used). Second, read the actual knowledge files and apply a single attribution test: Could the agent have reached the correct conclusion from its source materials?

• 如果材料中包含正确答案，但智能体仍然答错：说明是配方问题。
  If the materials contained the right answer but the agent still erred: recipe problem.
• 如果材料中没有包含正确答案：知识缺口。
  If the materials did not contain the right answer: knowledge gap.
• 如果专家们自己也对正确答案意见不一：存在歧义，标记出来供人工讨论。
  If experts themselves disagree on the right answer: ambiguity, flagged for human discussion.

编译：外科手术式多智能体编辑Compilation: Surgical Multi-Agent Edits

编译器将每个已诊断的问题转换为最小化的文件编辑。子代理并行分析影响，检查交叉引用、与现有位置的冲突、对令牌预算的影响、测试覆盖率以及重复风险。The compiler translates each diagnosed issue into minimal file edits. Sub-agents analyze impact in parallel, examining cross-references, conflicts with existing positions, token budget impact, test coverage, and duplication risk.

有两项设计选择使其值得信赖：Two design choices make this trustworthy:

独立对抗性审查。一个在全新上下文中运行、不了解改进理由的独立代理，只接收知识库的拟议差异。它的任务是找出诸如引入矛盾、破坏边缘情况或削弱既有立场等问题。由于它与提出方案的代理没有共享上下文，因此不会继承它们的盲点。Independent adversarial review. A separate agent, running in a fresh context with no knowledge of the improvement rationale, receives only the proposed diffs to the knowledge base. Its job is to find problems such as contradictions introduced, edge cases broken, or positions undermined. Because it shares no context with the proposing agents, it cannot inherit their blind spots.

确定性的结构验证。代码检查器会以编程方式捕获问题，包括悬空的交叉引用、文件大小预算违规、标识符冲突和依赖循环。这一层不是概率性的，而是通过或失败。Deterministic structural validation. A linter catches issues programmatically, dangling cross-references, file size budget violations, identifier collisions, and dependency cycles. This layer is not probabilistic. It passes or fails.

评估：证明修复有效Evaluation: Proving the Fix Works

每项拟议的更改都要经过两个阶段的验证：Every proposed change goes through a two-stage validation:

定向重放会让智能体在触发反馈的原始场景中运行。智能体并不知道自己正在接受测试。一个独立的评审器会根据原始专家反馈评估新的输出，同时不知道做了哪些改动。这种刻意的盲测设计可以避免确认偏误。如果定向重放失败，则会重试编译。Targeted replay runs the agent on the original scenario that triggered the feedback. The agent does not know it is being tested. A separate judge evaluates the new output against the original expert feedback without knowing what was changed. This deliberately blind design prevents confirmation bias. If targeted replay fails, compilation is retried.

回归测试会针对该领域运行多个基准测试，这些基准测试通常由问答对组成的结构化测试套件构成。对于可能存在多个正确答案的分析型领域，独立的 LLM 评判器会根据特定标准为每个测试用例判定通过或失败。代理会针对基准测试中的问题，在彼此独立的并行会话中运行，以检测性能回归。如果回归测试失败，则会使用更新后的提示重新进行编译；该提示会描述代理性能出现回归的地方，以及原始问题和尝试过的修复方案。Regression testing runs multiple benchmarks for that domain, which are usually structured test suites of Q&A pairs. For analytical domains whether there might be multiple correct answers, an independent LLM judge gives each test case a pass/fail based on certain criteria. The agent is run in parallel, independent sessions against the benchmark questions, and regressions in performance are detected. If regression testing fails, compilation is retried with an updated prompt describing where the agent regressed, along with the original issue and attempted fix.

落地与增值：复合式回报Landing and Enrichment: Compounding Returns

流水线的输出是一个包含完整审计轨迹的拉取请求（差异）。人类专家审查的是经过验证的修复方案，而不是调试未经处理的失败。一旦获得批准并合并（知识文件或配方随之更新），原始失败场景及其经过验证的正确答案就会自动添加到回归测试套件中。这意味着每次修复都会永久提高标准，而知识系统未来的更改必须保留刚刚纠正的行为。The output of the pipeline is a pull request (diff) with a complete audit trail. A human expert reviews a proven fix rather than debugging a raw failure. Once approved and landed (the knowledge file or recipe is updated), the original failing scenario and its validated correct answer are automatically added to the regression test suite. This means every fix permanently raises the bar and future changes to the knowledge system must preserve the behavior that was just corrected.

结果Results

经过为期六周的三个开发冲刺，系统实现了：After three development sprints spanning six weeks, the system achieved:

• 领域主题专家认为，智能体的输出几乎总是有用的，这相比早期版本有了显著改善；早期版本的输出经常需要进行大量返工。
  Domain SMEs rated agent outputs useful almost all the time, a significant improvement from early versions where outputs frequently required substantial rework.
• 将单项评估时间从数天缩短至数分钟。
  Days to minutes reduction in individual assessment time.
• 自动化自我改进，以过去需要完整工程冲刺周期的速度生成经过验证的知识编辑。
  Automated self-improvement producing validated knowledge edits at a rate that previously required full engineering sprints.
• 在各轮改进中实现零回归，并让每项修复都自动强化回归测试套件。
  Zero regressions across improvement cycles, with every fix automatically strengthening the regression suite.
• 领域专家一致反馈，该代理能够处理绝大多数分析工作，使他们可以专注于那些确实存在歧义、需要人类判断的案例。
  Domain experts consistently reported the agent handles the vast majority of the analytical work, allowing them to focus on the genuinely ambiguous cases that require human judgment.

应用此架构Applying This Architecture

我们为此构建的特定领域需要将数十个来源（包括内部立场和外部材料）综合成按风险加权的评估。但该架构与具体领域无关，适用于以下场景：The specific domain we built this for required synthesizing dozens of sources (both internal positions and external material) into risk-weighted assessments. But the architecture is domain-independent. It applies wherever:

• 专业知识以专家头脑中的部落知识形式存在。
  Specialist knowledge lives as tribal knowledge in experts’ heads.
• 各项评估之间的一致性很重要。
  Consistency across assessments matters.
• 工作量超出了现有专家的处理能力。
  The volume of work exceeds available expert capacity.
• 现成的 LLM 无法生成充分的分析。
  Off-the-shelf LLMs produce inadequate analysis.

适合这一模式的具体领域包括监管合规、协议遵循、金融风险评估、安全审查、工程标准合规以及采购评估。它们的共同点在于，组织需要的是具备真正机构专业知识的 AI 系统，而不仅仅是掌握一般知识的系统。Concrete domains where this pattern fits include regulatory compliance, protocol adherence, financial risk assessment, security review, engineering standards compliance, and procurement evaluation. The common thread is that organizations need AI systems with genuine institutional expertise, not just general knowledge.

采用此架构的要求如下：The requirements for adopting this architecture are:

1. 一个具有明确文件边界、交叉引用和依赖关系图的结构化知识系统（该领域的组织型第二大脑）。
   A structured knowledge system with explicit file boundaries, cross-references, and a dependency graph (the organizational second brain for the domain).
2. 一种将领域知识与分析方法（配方）分离的程序化层。
   A procedural layer that separates domain knowledge from analytical methodology (recipes).
3. 一个随着每轮改进不断扩展的自动化评估套件。
   An automated evaluation suite that grows with each improvement cycle.
4. 根据该领域的风险容忍度校准的人机协同检查点。
   Human-in-the-loop checkpoints calibrated to the domain’s risk tolerance.

更深层的原则很简单：将复杂性保留在既便于人类阅读、又便于智能体读取的文本文件中，而不是写入经过微调的模型权重。每一项改进都是文本编辑，领域专家可以在 30 秒内完成审阅。每次变更都有版本控制、可进行差异比较，并且可以撤销。编译流水线很复杂，但其输出始终是透明的。The deeper principle is straightforward: keep the complexity in text files that are readable by both humans and agents, rather than fine-tuned model weights. Every improvement is a text edit that a domain expert can review in 30 seconds. Every change is version-controlled, diffable, and reversible. The compilation pipeline is sophisticated, but its outputs are always transparent.

目标是构建一个让专家的努力得以永久复利的系统。每一次专家互动都会让系统变得更好。每一次纠正都会作为经过验证的改进持续保留。组织的集体知识不再被困在个人手中，而是能够稳定地、大规模地提供给每一位需要它的人。The goal is a system where expert effort compounds permanently. Every expert interaction makes the system better. Every correction persists as a verified improvement. The organization’s collective knowledge stops being trapped in individuals and starts being available, consistently and at scale, to everyone who needs it.

致谢Acknowledgements

作者谨向以下人员为本系统开发所作出的贡献表示感谢，他们在该系统的开发过程中发挥了至关重要的作用。特别是，我们要向以下人员致以特别感谢（按姓氏字母顺序）：Cecilia Baek、Philipp Kaufold、Cat Hughes、Suzanne Leijten、Michael Marcusa、Jordi Mola、Timothy Neo、Elliott Prentiss、Laia Reyes、John Ross、Julio Santil、Taylor Wilson Thomas、Mansi Tripathi、Nikhil Shanbhag、Madeleine Vos 和 Jackie Zajac。The authors would like to express our gratitude to the contributions of the following people, who have played a crucial role in developing this system. In particular, we would like to extend special thanks to (in last name alphabetical order): Cecilia Baek, Philipp Kaufold, Cat Hughes, Suzanne Leijten, Michael Marcusa, Jordi Mola, Timothy Neo, Elliott Prentiss, Laia Reyes, John Ross, Julio Santil, Taylor Wilson Thomas, Mansi Tripathi, Nikhil Shanbhag, Madeleine Vos, and Jackie Zajac.