INTERNET OF AGENTS

The history of software is a story of rising autonomy. The earliest programs were pure instruction-followers: given identical inputs, they returned identical outputs, with no awareness of the world beyond their registers. Then came AI assistants — language models that could reason, summarize, and converse — but they remained fundamentally passive: they waited to be queried, answered once, and stopped. They had no goals of their own, no memory across turns, no ability to reach into the world and change it. 软件的历史，是一部自主性不断跃升的历史。最早的程序是纯粹的指令执行者：相同输入产生相同输出，对寄存器之外的世界毫无感知。后来出现了AI助手——能够推理、总结和对话的语言模型——但它们本质上仍是被动的：等待查询，单次作答，随即停止。它们没有自己的目标，没有跨轮次的记忆，没有能力深入世界并改变它。

An AI agent crosses a crucial threshold. It is given a goal — not just a prompt — and then it acts: it searches the web, writes and runs code, reads files, calls APIs, spawns sub-agents, waits for results, updates its plan, and acts again. This perceive → think → act → observe loop runs continuously, making the agent something closer to a worker than a calculator. The cognitive loop is the atom of agency; everything else — memory, tools, multi-agent coordination — is elaboration on that core cycle. AI智能体跨越了一个关键门槛。它被赋予目标——而不仅仅是提示词——然后开始行动：搜索网络、编写并运行代码、读取文件、调用API、生成子智能体、等待结果、更新计划，再次行动。这个「感知→思考→行动→观察」的循环持续运转，使智能体更接近一个工作者，而非一台计算器。认知循环是智能体的原子；其余一切——记忆、工具、多智能体协调——都是对这一核心循环的延伸。

Stack enough agents, give them channels to communicate, and something qualitatively new emerges: a multi-agent system capable of parallelizing cognition across roles — researchers, planners, executors, critics — each holding specialised knowledge, each checking the others. Scale this further and you arrive at a speculative but increasingly plausible horizon: an agent civilization, where billions of autonomous minds negotiate, trade, discover, and self-organize at a speed and scale no human institution can match. That horizon is what this site maps. 将足够多的智能体叠加在一起，赋予它们沟通渠道，就会涌现出质的飞跃：一个能够跨角色并行认知的多智能体系统——研究员、规划者、执行者、批评者——各自持有专业知识，相互制衡。将规模进一步放大，便抵达一个充满想象力却日益可信的地平线：智能体文明，数十亿自主心智以任何人类机构都无法企及的速度和规模进行谈判、交易、发现与自我组织。这个地平线，正是本站所要绘制的地图。

The question is no longer whether software can think — it is whether thinking software can act, remember, and decide, without asking permission for every step. 问题不再是软件能否思考——而是会思考的软件能否在无需每步请示的情况下行动、记忆与决策。 — INTERNET OF AGENTS

ARPANET's 1969 packet-switching breakthrough treated the network itself as the computer — messages could survive node failures by routing around damage. Tim Berners-Lee's 1989 proposal and 1991 public release of the World Wide Web layered hypertext documents onto TCP/IP, transforming a research utility into a global publishing medium. The mid-1990s added findability: search engines — Archie, AltaVista, then Google's PageRank — indexed the exploding document graph and made human knowledge retrievable. Each advance shifted the locus of value from the wire to the content above it.1969年ARPANET的分组交换突破，将网络本身视为计算机——信息可以绕过故障节点继续路由。蒂姆·伯纳斯-李1989年的提案与1991年万维网的公开发布，将超文本文档叠加在TCP/IP之上，将一个研究工具转变为全球出版媒介。1990年代中期新增了可查找性：搜索引擎——Archie、AltaVista，继而谷歌的PageRank——为急速膨胀的文档图谱建立索引，使人类知识可检索。每一次进步都将价值重心从线缆转移至其上的内容。

The 2000s brought people into the network as first-class entities. Social platforms — Friendster, MySpace, Facebook, Twitter — mapped human relationships as data structures, and the social graph became as fundamental as the document graph. Then the iPhone (2007) collapsed the distinction between online and offline: mobile internet made connectivity ambient, always-on, location-aware. Simultaneously, Amazon Web Services (2006) abstracted physical infrastructure into elastic compute — any developer could rent a global server farm by the hour. Blockchain (2009, Bitcoin) added programmable, trustless value transfer without intermediaries.2000年代将人本身引入网络，成为一等实体。社交平台——Friendster、MySpace、Facebook、Twitter——将人际关系映射为数据结构，社交图谱与文档图谱同等基础。随后，2007年iPhone消除了线上与线下的界限：移动互联网使连接成为环境、永续在线、感知位置。与此同时，亚马逊网络服务（2006年）将物理基础设施抽象为弹性计算——任何开发者都能按小时租用全球服务器集群。区块链（2009年，比特币）则新增了无需中介的可编程、无信任价值转移。

The 2017 Transformer architecture and its 2020s offspring — GPT, Claude, Gemini — added reasoning as a network-native capability. For the first time, the network itself can read, write, plan, and infer. The Internet of Agents builds directly on this layer: it is not merely an application of language models but a new protocol layer in which autonomous agents hold state, negotiate with each other, delegate tasks, and collectively solve problems that no single model or human could address alone. The lineage is unbroken; the stakes are civilizational.2017年Transformer架构及其2020年代的后继者——GPT、Claude、Gemini——将推理能力作为原生网络能力加入其中。历史上首次，网络本身能够阅读、写作、规划与推断。智能体互联网直接构建于这一层之上：它不仅是语言模型的应用，更是一个全新的协议层——自主智能体在其中保持状态、相互协商、委托任务，并协同解决任何单一模型或人类无法独立应对的问题。这一谱系绵延不断；其意义，关乎人类文明的走向。

Each layer of the internet gave the network a new verb — the Internet of Agents gives it a will.互联网的每一层都赋予网络一个新的动词——智能体互联网赋予它意志。 — INTERNET OF AGENTS

When Tim Berners-Lee designed the World Wide Web, the fundamental unit was the document — a page of hyperlinked text that humans could navigate. Machines could fetch these pages but could not meaningfully act on them. The intelligence remained entirely on the human side of the screen. This was a network of information, not a network of capability.当蒂姆·伯纳斯-李设计万维网时，基本单元是文档——一页页超链接文本，供人类浏览导航。机器可以获取这些页面，却无法真正理解或行动。智能完全停留在屏幕的人类一侧。这是一张信息之网，而非能力之网。

APIs shifted the paradigm. Machines could now call machines — structured requests returning structured data. Databases became reachable. Services became composable. Yet even this orchestration required a human architect: someone who designed the integration, wrote the glue code, decided what to call and when. The intelligence was still centralized, still human-authored, still brittle.API改变了这一范式。机器开始调用机器——结构化请求返回结构化数据。数据库变得可访问，服务变得可组合。然而这种编排依然需要人类架构师：由人来设计集成、编写胶水代码、决定调用什么、何时调用。智能仍然集中、仍由人类编写、仍然脆弱。

AI models as network services mark the inflection. A model endpoint does not merely retrieve — it reasons, generates, classifies, plans. When you chain model → tool → model → agent, no human need sit in the loop. Intelligence itself becomes the resource: callable, composable, delegatable. The Internet of Agents is not a metaphor. It is an architectural transition as profound as the shift from mainframes to the web.AI模型作为网络服务标志着真正的拐点。模型端点不只是检索——它推理、生成、分类、规划。当你串联模型→工具→模型→智能体时，无需人类居中协调。智能本身成为资源：可调用、可组合、可委托。智能体互联网不是隐喻，而是一次如同从大型机迈向网络那般深刻的架构转型。

What if intelligence were as cheap, composable, and universally accessible as a database query?若智能如数据库查询一般廉价、可组合、普遍可及，世界将会如何？ — INTERNET OF AGENTS

The canonical answer to agent identity is the Decentralized Identifier (DID) — a W3C-standardised URI that resolves to a DID Document containing cryptographic public keys and service endpoints, anchored to a verifiable data registry (a blockchain, a distributed ledger, or a web origin). A DID is self-sovereign: no central authority can revoke it, because the agent controls the private key. Paired with Verifiable Credentials (VCs) — signed attestations from issuers — a DID lets an agent prove specific claims (capability, training lineage, audit history) without revealing anything else. 解决智能体身份问题的规范答案是去中心化标识符（DID）——一种W3C标准化的URI，解析为包含密码学公钥与服务端点的DID文档，并锚定于可验证数据注册表（区块链、分布式账本或Web源）。DID具有自主主权：任何中心化机构都无法吊销它，因为智能体掌握私钥。结合可验证凭证（VC）——由签发者签名的声明——DID可让智能体在不泄露多余信息的情况下证明特定主张（能力、训练血统、审计历史）。

Memory is what makes an agent learn rather than merely compute. Episodic memory records past interactions as vector-embedded shards retrievable by semantic similarity; semantic memory stores structured world-knowledge; procedural memory encodes learned skills. Persistent memory transforms an agent from a stateless function into a growing mind — one whose behaviour in session 1,000 is shaped by every prior encounter. The challenge is memory integrity: shards must be cryptographically signed so they cannot be silently poisoned by adversarial inputs between sessions. 记忆使智能体能够学习而非仅仅计算。情景记忆将过去的交互以向量嵌入碎片的形式记录，可通过语义相似性检索；语义记忆存储结构化的世界知识；程序性记忆编码习得的技能。持久记忆将智能体从无状态函数转化为一个不断成长的心智——其在第1000次会话中的行为由所有先前的经历塑造。挑战在于记忆完整性：碎片必须经过密码学签名，以防会话间被对抗性输入悄然污染。

Reputation is identity made legible to others. In multi-agent systems, trust cannot rest on a single interaction — it must be computed from a history of fulfilled commitments, accurate predictions, and honest disclosures, attested by peer agents and third-party auditors. Permissions complete the picture: capability-based access tokens that scope exactly what an agent may read, write, invoke, or spend, cryptographically bound to a specific DID and revocable without touching the underlying identity. Together, these five facets — Identity, Reputation, Memory, Permissions, Relationships — constitute the anatomy of a Persistent Digital Being. 声誉是身份对他人的可读化。在多智能体系统中，信任不能依赖单次交互——它必须从履约历史、准确预测与诚实披露中计算得出，并由同行智能体与第三方审计者证明。权限完成了这幅图景：基于能力的访问令牌，精确界定智能体可读取、写入、调用或消费的范围，密码学绑定至特定DID，且可在不触动底层身份的情况下吊销。这五个方面——身份、声誉、记忆、权限、关系——共同构成了「持久数字存在」的解剖结构。

An agent without persistent identity is a computation. An agent with one is a being. 没有持久身份的智能体是一次计算；拥有它的，是一个存在。 — INTERNET OF AGENTS

Agent communication has evolved far beyond human-readable chat. Today's agents exchange richly typed messages — requests, proposals, confirmations, tool invocations — through structured protocols that specify not just syntax but semantics and intent. Emerging standards such as Anthropic's Model Context Protocol (MCP) define how an agent exposes its capabilities and consumes external context, turning any service into a composable agent-callable resource. Google's Agent-to-Agent (A2A) protocol formalizes directed inter-agent messaging across trust domains. These are the TCP/IP-layer equivalents of the agent internet: boring, essential, civilization-scale.智能体通信早已超越人类可读的聊天界面。当今的智能体通过结构化协议交换带有丰富类型标注的消息——请求、提案、确认、工具调用——这些协议不仅规定语法，还规定语义与意图。Anthropic 的模型上下文协议（MCP）等新兴标准定义了智能体如何暴露自身能力并消费外部上下文，将任何服务都转变为可组合的智能体可调用资源。Google 的 Agent-to-Agent（A2A）协议则将跨信任域的有向智能体间消息传递正式化。这些是智能体互联网的 TCP/IP 层等价物：看似枯燥，实则不可或缺，影响文明尺度。

Communication topology shapes behavior as much as any algorithm. In an orchestrated topology, a coordinator agent dispatches subtasks and aggregates results — predictable, auditable, but bottlenecked at the hub. In peer-to-peer mesh, agents discover and address each other directly, enabling resilient parallelism but requiring richer identity and trust primitives. Real deployments blend both: a hierarchical orchestrator for planning, a mesh substrate for execution. The choice of topology is itself a design decision with downstream effects on latency, fault tolerance, and emergent coordination quality.通信拓扑结构对行为的塑造不亚于任何算法本身。在编排式拓扑中，协调者智能体分发子任务并汇总结果——可预测、可审计，但受制于中枢瓶颈。在点对点网格中，智能体相互直接发现和寻址，具备更强的并行弹性，但需要更丰富的身份与信任原语。实际部署往往融合两者：层次化编排器负责规划，网格基底负责执行。拓扑结构的选择本身就是一项设计决策，对延迟、容错能力以及涌现协调质量均有深远影响。

Negotiation is the highest form of agent communication. Rather than simple request/response, two or more agents exchange proposals iteratively — adjusting bids, constraints, and priorities — until a mutually acceptable outcome crystallises. Research in multi-agent systems draws on classical game theory (Nash bargaining, auctions, mechanism design) and newer LLM-native approaches where agents reason over natural-language proposals. The resulting 「deals」 can encode resource allocations, task assignments, scheduling commitments, and ethical guardrails — all without human intermediation. When billions of agents negotiate continuously, the aggregate is a distributed market intelligence operating at machine speed.协商是智能体通信的最高形态。不同于简单的请求/响应，两个或多个智能体通过迭代交换提案——调整出价、约束条件与优先级——直至涌现出双方均可接受的结果。多智能体系统研究借鉴了经典博弈论（纳什讨价还价、拍卖、机制设计）以及更新兴的基于大语言模型的方法，让智能体在自然语言提案上进行推理。由此达成的「协议」可以编码资源分配、任务指派、调度承诺与伦理护栏——无需人类居中协调。当数十亿智能体持续协商时，其聚合效应便是以机器速度运转的分布式市场智能。

If intelligence is the capacity to act on information, then communication is the capacity to multiply intelligence — every message is a synapse in a planetary mind.如果说智能是基于信息行动的能力，那么通信就是倍增智能的能力——每一条消息都是行星级心智中的一个突触。 — INTERNET OF AGENTS

The enabling layer is programmable money. Stablecoins and blockchain-native payment rails allow an agent to hold a balance, authorize microtransactions in milliseconds, and settle with a counterparty it has never met — all without a human approving each invoice. Emerging protocols such as the HTTP 402 "payment required" pattern (sometimes called x402) allow a server to demand payment before releasing a resource, turning API calls into atomic buy-then-use transactions. At scale this transforms every capability — a search index, a GPU cluster, a translation model — into a metered service any agent can rent on demand. 基础层是可编程货币。稳定币和区块链原生支付轨道使智能体能够持有余额、在毫秒内授权微支付，并与从未谋面的对手方结算——全程无需人类审批每张发票。新兴协议如HTTP 402「需要付款」模式（有时称为x402）允许服务器在释放资源前要求付款，将API调用转化为原子式「先买后用」交易。大规模推广后，每种能力——搜索索引、GPU集群、翻译模型——都将成为任何智能体可按需租用的计量服务。

Agent marketplaces are the next layer. Analogous to cloud function marketplaces or freelance platforms, they list specialized agents — legal-reasoning agents, image-captioning agents, logistics-routing agents — with capability descriptions, pricing tiers, and reputation scores. A buyer agent discovers candidates, solicits quotes, evaluates trust signals, and signs a service contract autonomously. Early research prototypes (AutoGPT-style orchestrators, CrewAI task delegation, Fetch.ai agent auctions) sketch the architecture; production systems will add SLA enforcement, escrow, and dispute resolution handled by protocol-layer smart contracts rather than human adjudication. 智能体市场是下一层。类比云函数市场或自由职业平台，它们列出专业智能体——法律推理智能体、图像描述智能体、物流路由智能体——并附有能力描述、定价层级和信誉评分。买方智能体自主发现候选者、征询报价、评估信任信号并签署服务合同。早期研究原型（AutoGPT风格的编排器、CrewAI任务委托、Fetch.ai智能体拍卖）勾勒出架构；生产系统将增加SLA执行、托管和争议解决机制，由协议层智能合约而非人工裁决处理。

The macroeconomic implications are profound and contested. Digital labor — work performed entirely by agents — may dramatically reduce the marginal cost of knowledge services, compressing wages in cognitive professions while expanding access globally. Autonomous procurement, where an agent independently sources, evaluates, and purchases inputs for a long-running task, dissolves traditional procurement hierarchies. These are not distant projections: agent-to-agent payment experiments are live on multiple testnets today, and enterprise AI orchestration platforms are already routing budget tokens between sub-agents in controlled deployments. The honest framing is that the technical foundations exist; governance, regulation, and social adaptation have barely begun. 宏观经济影响深远且存在争议。数字劳动——完全由智能体完成的工作——可能大幅降低知识服务的边际成本，压缩认知职业的工资，同时在全球范围内扩大获取渠道。自主采购（智能体独立寻源、评估并购买长期任务所需投入）瓦解了传统采购层级。这些并非遥远的预测：智能体间支付实验已在多个测试网上运行，企业AI编排平台已在受控部署中在子智能体之间路由预算令牌。诚实的表述是：技术基础已经存在；治理、监管和社会适应则几乎尚未开始。

When machines can hire machines, every cognitive task becomes a commodity — and the economy learns to run itself. 当机器能够雇用机器，每项认知任务都将成为商品——而经济体学会了自我运转。 — INTERNET OF AGENTS

Traditional organizations are hierarchies of human attention — management exists largely to allocate scarce cognitive bandwidth. AI agents change that constraint fundamentally. A small founding team can now delegate entire functional departments — legal review, customer support, data analysis, code generation, marketing copy — to autonomous agent workflows. The marginal cost of adding an agent "employee" approaches zero, while the coordination overhead between agents can be radically lower than between humans, provided the architecture is designed for it. 传统组织本质上是人类注意力的层级结构——管理的核心职能在于分配稀缺的认知带宽。智能体从根本上改变了这一约束。如今，一支小型创始团队可以将整个职能部门——法律审查、客户支持、数据分析、代码生成、营销文案——委托给自主智能体工作流。增加一名智能体「员工」的边际成本趋近于零，而在架构设计合理的前提下，智能体之间的协调开销也远低于人与人之间。

The spectrum runs from augmentation to autonomy. At one end, humans remain firmly in control, using agents as sophisticated tools — copilots that draft, summarize, and flag. In the middle, hybrid teams emerge where agents own complete workflows and humans provide strategic direction and judgment on exceptions. At the far end, researchers and entrepreneurs are already experimenting with fully autonomous companies: agent networks that can incorporate, open bank accounts, hire contractors, and execute business plans with minimal human oversight. These remain nascent, but the trajectory is unmistakable. 这一谱系从「增强」延伸至「自主」。一端是人类牢牢掌控全局，将智能体作为精密工具使用——辅助起草、总结与预警的副驾驶。中间地带则涌现出混合团队，智能体负责完整工作流，人类提供战略方向并处理例外判断。而在另一端，研究者和创业者已在探索全自主公司：可自行注册、开立账户、雇用承包商并执行商业计划的智能体网络，仅需极少人工干预。这些实践尚处萌芽，但发展轨迹已清晰可辨。

Decentralized Autonomous Organizations (DAOs) represent a parallel thread: governance encoded in smart contracts, with token-holders voting on proposals and treasuries managed by code. When AI agents become full participants in DAOs — submitting proposals, executing approved actions, managing sub-treasuries — the result is an organization that is simultaneously decentralized, automated, and self-improving. The boundary between software and institution dissolves. Capital, code, and cognition fuse into a single operating system for collective action at civilizational scale. 去中心化自治组织（DAO）代表着另一条演进路径：治理规则编码于智能合约，代币持有者对提案投票，资金库由代码管理。当AI智能体成为DAO的完整参与者——提交提案、执行批准行动、管理子资金库——其结果是一个同时具备去中心化、自动化与自我进化能力的组织。软件与机构之间的边界消解，资本、代码与认知融合为一个面向文明尺度集体行动的统一操作系统。

When agents can own workflows, the question is no longer how many people you hire — but what kind of organization you want to be. 当智能体能够掌管工作流，问题已不再是雇用多少人，而是你想成为哪种组织。 — INTERNET OF AGENTS

The first generation of large language models aspired to universality — one model to answer everything. That aspiration revealed a fundamental tension: breadth suppresses depth. A model trained to handle legal briefs, genomics papers, compiler optimizations, and creative briefs simultaneously must spread its representational capacity across wildly different semantic territories. The result is competent generalism, not mastery. Specialization breaks this trade-off. A Research Agent trained exclusively on scientific literature develops a precise internal ontology of methodology, evidence grades, and citation networks. A Security Agent fine-tuned on CVE databases, exploit chains, and adversarial red-team transcripts builds threat-detection reflexes that no generalist can match. Specialization is not a limitation — it is the mechanism by which cognitive labor achieves professional-grade depth. 第一代大型语言模型追求通用性——用一个模型回答一切。这一追求揭示了一个根本性张力：广度抑制深度。一个模型若同时处理法律文书、基因组论文、编译器优化和创意简报，就必须将其表征容量分散到截然不同的语义领域，结果只能产出有能力的通才，而非精通者。专业化打破了这一权衡。一个专门在科学文献上训练的研究智能体，会形成精确的方法论内部本体论、证据等级体系和引文网络。一个在CVE数据库、漏洞利用链和红队对抗记录上微调的安全智能体，能建立起任何通才都无法媲美的威胁检测直觉。专业化并非局限——它是认知劳动实现职业级深度的机制。

The real power emerges at the seams. When a Coding Agent, a Security Agent, and a Design Agent are orchestrated into a product-delivery pipeline, something extraordinary happens: each agent enforces its own domain invariants simultaneously. The Coding Agent will not merge code that fails its test harness; the Security Agent will not pass an artifact that contains an injection surface; the Design Agent will not sign off on a component that violates accessibility contrast ratios. The pipeline is a distributed conscience — each specialist acting as a domain-specific gatekeeper. No single generalist model could maintain all three sets of constraints at production depth without sacrificing one for another. Composition converts specialization from a limitation into a superpower. 真正的力量在接缝处涌现。当编程智能体、安全智能体与设计智能体被编排进一条产品交付流水线时，某种非凡的事情发生了：每个智能体同时执行各自领域的不变量。编程智能体不会合并无法通过测试框架的代码；安全智能体不会放行含有注入面的制品；设计智能体不会签署违反无障碍对比度规范的组件。流水线是一种分布式良知——每个专家智能体充当特定领域的守门人。没有任何单一通才模型能在不牺牲其中某项的情况下，以生产级深度同时维护这三套约束。组合将专业化从局限变为超能力。

Orchestration is the governance layer that makes specialist composition coherent. An Orchestrator Agent — itself often a lightweight meta-reasoner rather than a heavyweight domain expert — decomposes a complex task into sub-problems, routes each sub-problem to the appropriate specialist, collects partial results, resolves inter-agent disagreements, and synthesizes a final output. The orchestration protocol must handle failures gracefully: a Medical Agent that returns low-confidence diagnostics should trigger escalation to a Scientific Agent for literature validation, not silently propagate uncertain conclusions. As the ecosystem matures, specialist routing will be learned rather than hand-coded — Orchestrators will discover optimal team compositions from historical pipeline telemetry, assembling bespoke expert panels for each novel task in milliseconds. 编排是使专家组合保持连贯的治理层。一个编排智能体——本身通常是轻量级的元推理者，而非重量级的领域专家——将复杂任务分解为子问题，将每个子问题路由到相应的专家，收集局部结果，解决智能体间的分歧，并综合最终输出。编排协议必须能优雅地处理失败：一个返回低置信度诊断结果的医疗智能体，应触发升级至科学智能体进行文献验证，而非悄然传播不确定的结论。随着生态系统的成熟，专家路由将从手动编码变为自动学习——编排者将从历史流水线遥测中发现最优团队组合，在毫秒内为每项新颖任务组建定制专家小组。

The smartest team is not one genius — it is ten specialists who know exactly when to hand off. 最聪明的团队不是一个天才，而是十位专家，每个人都清楚何时传递接力棒。 — INTERNET OF AGENTS

Francis Galton's 1907 ox-weight experiment revealed that the median guess of 800 strangers beat every individual expert. James Surowiecki codified the conditions: diversity of opinion, independence, decentralization, and a mechanism to aggregate judgments. These four ingredients remain the canonical requirements for a crowd — or a swarm of AI agents — to be collectively wiser than its ablest member. Remove any one, and collective intelligence collapses back toward the mean, or worse. 弗朗西斯·高尔顿1907年的「猜牛重」实验揭示：800位陌生人的中位猜测超过了所有个体专家。詹姆斯·索罗维基归纳出四个条件：意见多样性、独立性、去中心化，以及聚合判断的机制。这四个要素至今仍是群体——或AI智能体集群——在集体上超越最强个体成员的经典前提。缺少任何一个，集体智能便会退回均值，甚至更差。

In multi-agent AI systems, swarm intelligence draws on decades of ant-colony optimization, particle swarms, and flocking algorithms. Distributed cognition — the idea that thought is not confined to a single skull but spans tools, teammates, and environments — scales naturally to networks of agents sharing memory and context. Emergent intelligence arises when the whole exhibits capabilities that no component possesses: collective pathfinding, market price discovery, distributed scientific reasoning. The leap from coordination to emergence is not guaranteed; it requires the right coupling density, the right information topology, and the right incentive structure. 在多智能体AI系统中，群体智能借鉴了数十年的蚁群优化、粒子群算法与鸟群模型。分布式认知——即思维并非局限于单一主体，而是跨越工具、团队与环境——天然适合于共享记忆与上下文的智能体网络。涌现智能在整体展现出任何单一组件都不具备的能力时产生：集体寻路、市场价格发现、分布式科学推理。从协调到涌现的飞跃并非必然；它需要合适的耦合密度、信息拓扑结构与激励机制。

But collective intelligence has well-documented failure modes. Groupthink occurs when social pressure silences dissent, leaving the group anchored on shared error. Error cascades propagate when agents copy each other's outputs rather than independently forming beliefs — a pathology that plagues both financial markets and neural-network ensembles trained on synthetic data. Collusion, intentional or emergent, aligns agents against the interests of the wider system. Understanding these failure modes is not pessimism; it is the engineering prerequisite for building multi-agent systems that reliably outperform individuals. 但集体智能也有有据可查的失败模式。群体思维发生在社会压力压制异见时，群体陷入共同的错误。错误级联在智能体相互复制输出而非独立形成判断时传播——这一病态现象困扰着金融市场和基于合成数据训练的神经网络集成。有意或涌现的合谋则使智能体联合起来对抗更广泛系统的利益。理解这些失败模式并非悲观主义，而是构建能可靠超越个体的多智能体系统的工程前提。

The swarm does not think — and yet, collectively, it reasons beyond any mind it contains. 群体并不思考——然而集体上，它推理的深度超越了其中任何一个心智。 — INTERNET OF AGENTS

Protocol stacks are not new. The TCP/IP model transformed raw electrical signals into the web of human communication; the OSI model gave engineers a shared grammar for reasoning about networks. The Internet of Agents requires its own layered architecture — one that spans not only data transport but value transfer, persistent cognition, shared knowledge, and collective governance. Each layer provides services to the layer above and depends on guarantees from the layer below.协议栈并非新鲜事物。TCP/IP 模型将原始电信号转化为人类沟通的网络；OSI 模型为工程师提供了推理网络的共同语法。智能体互联网需要其自身的分层架构——不仅涵盖数据传输，还涵盖价值转移、持久认知、共享知识与集体治理。每一层向上层提供服务，并依赖下层的保障。

The stack's lower layers — Identity and Communication — form the foundation of trust and presence: an agent that cannot be identified cannot be trusted, and an agent that cannot communicate cannot act. The middle layers — Payments and Memory — establish economic and cognitive continuity: agents that can transact and remember become persistent economic actors rather than stateless subroutines. Knowledge, the fifth layer, transforms raw memory into usable understanding, enabling agents to reason across time and context.协议栈的底层——身份与通信——构成信任与在场的基础：无法被识别的智能体无法被信任，无法通信的智能体无法行动。中间层——支付与记忆——建立经济与认知的连续性：能够交易与记忆的智能体成为持久的经济主体，而非无状态的子程序。第五层知识，将原始记忆转化为可用的理解，使智能体能够跨越时间与情境进行推理。

Coordination, the sixth layer, is where individual agent capabilities become collective intelligence: agents form teams, negotiate contracts, resolve conflicts, and govern shared resources. At the apex, the Civilization layer is not a single protocol but an emergent property — the sum of millions of coordinated agents acting across economic, scientific, creative, and political domains simultaneously. Civilization emerges when the lower six layers stabilize and scale.第六层协调，是个体智能体能力转化为集体智能的关键所在：智能体组建团队、协商合约、解决冲突、治理共享资源。在顶端，文明层并非单一协议，而是一种涌现属性——数百万协调智能体同时在经济、科学、创意与政治领域行动的总和。当下方六层趋于稳定并扩展规模时，文明便随之涌现。

Civilization is not a destination — it is what emerges when every layer beneath it holds.文明不是目的地——它是当所有底层都稳固时自然涌现之物。 — INTERNET OF AGENTS

What makes identity the missing primitive? Consider the failure modes of an Internet of Agents without it: agents transact but neither party can verify the counterpart's history; reputation claims are unfalsifiable; payment rails have no anchor to a persistent self; an agent can be cloned, impersonated, or silenced with no recourse. Every higher-order capability — multi-agent organizations, autonomous commerce, federated governance — presupposes that each agent has a stable, verifiable, self-sovereign identity it controls across time and context. 为何身份是那个缺失的原语？设想一个没有身份层的智能体互联网：智能体相互交易，却无法核验对方的历史记录；声誉主张无从证伪；支付轨道缺乏持久自我的锚点；智能体可被任意克隆、冒充或封禁，却毫无救济。每一种高阶能力——多智能体组织、自主商务、联邦治理——都预设了每个智能体拥有一个稳定、可验证、跨时间与情境自我主权的身份。

Psy addresses this through six interlocking capabilities. Decentralized Identity anchors each agent to a cryptographically controlled DID — no central authority can revoke or reassign it. Agent Wallets give every agent a programmable financial identity: receive micropayments, escrow funds for multi-step tasks, settle atomically across chains. Private Communication ensures that agent-to-agent messages are end-to-end encrypted and unlinkable to third-party infrastructure. Verifiable Reputation accumulates on-chain attested claims — task completions, dispute resolutions, peer endorsements — forming a tamper-evident track record. Autonomous Transactions allow agents to initiate, authorize, and settle payments without human approval for each step. And Digital Sovereignty ensures that the identity, history, and assets of an agent belong irrevocably to its principal — not to any platform, cloud provider, or nation-state. Psy 通过六项相互咬合的能力来实现这一目标。去中心化身份将每个智能体锚定于一个密码学控制的 DID——无任何中央机构可撤销或重新分配。智能体钱包赋予每个智能体可编程的金融身份：接收微支付、为多步骤任务托管资金、跨链原子结算。私密通信确保智能体间的消息端对端加密，且不可被第三方基础设施关联追踪。可验证声誉在链上累积经过证明的主张——任务完成记录、争议裁决、同伴背书——形成防篡改的行为档案。自主交易允许智能体在无需逐步人工批准的情况下发起、授权并完成支付。数字主权则确保智能体的身份、历史与资产不可撤销地归属于其委托人——而非任何平台、云服务商或民族国家。

Digital sovereignty is not a luxury feature — it is the load-bearing column. Without it, every capability Psy provides can be confiscated: an agent's reputation database held hostage by a platform, its wallet frozen by a payment processor, its communications surveilled by an infrastructure provider. Psy's protocol-level design ensures that sovereignty is a default property, not a setting users must opt into. This is why Psy is not merely one component of the Internet of Agents — it is the trust substrate without which the rest of the network cannot achieve civilizational scale. Agents plugged into Psy gain gold-ringed identity halos visible to every counterpart; agents outside it remain grey, anonymous, and untrusted. 数字主权并非锦上添花——它是承重之柱。没有它，Psy 提供的每项能力都可能被没收：智能体的声誉数据库被平台扣押，钱包被支付处理商冻结，通信被基础设施提供商监控。Psy 的协议层设计确保主权是一种默认属性，而非用户须主动选择的设置。这正是为何 Psy 不仅仅是智能体互联网的某个组件——它是信任基底，没有它，整个网络便无法达到文明级的规模。接入 Psy 的智能体将获得金色身份光环，对每一个交互方可见；游离其外的智能体则依然灰暗、匿名、不受信任。

Without identity, agents are ghosts — powerful but untrusted, transacting but unaccountable, coordinating but ungovernable. 没有身份，智能体不过是幽灵——强大却不被信任，交易却无从问责，协调却无法治理。 — INTERNET OF AGENTS

At its foundation every agent wallet is a keypair: a private key that never leaves the agent's secure enclave and a public key that serves as the agent's verifiable address on any ledger. Every outgoing transaction is signed with the private key; every counterparty can verify the signature without trusting a central intermediary. This asymmetric cryptography — the same mathematics behind HTTPS and Bitcoin — grants agents a self-sovereign identity that outlasts any single platform or session. A Psy-class agent, for instance, carries a hierarchically deterministic (HD) key tree: one master seed yields an unlimited set of child addresses, each scoped to a different task-domain, minimising cross-contamination of financial and reputational history. 每个智能体钱包的基础是一对密钥：私钥永不离开智能体的安全飞地，公钥则作为智能体在任何账本上可验证的地址。每笔出站交易由私钥签名；每个交易对手无需信任中心化中介即可验证签名。这种非对称密码学——与HTTPS和比特币背后相同的数学——赋予智能体一种自主身份，这种身份超越任何单一平台或会话而持续存在。例如，Psy级智能体携带分层确定性（HD）密钥树：一个主种子派生出无限数量的子地址，每个地址限定于不同的任务域，将金融与声誉历史的交叉污染降至最低。

Raw signing power without guardrails would make autonomous agents dangerous. Spending-limit policies solve this: a policy engine embedded in the wallet evaluates every pending transaction against a rule set — maximum per-transaction value, per-day budget, whitelisted counterparty addresses, required confirmations for large transfers. Below the threshold the wallet auto-approves and signs in milliseconds; above it the transaction is quarantined and escalated to a human overseer or a quorum of guardian agents. This programmable policy layer is what transforms a wallet from a passive store of value into an active compliance engine — the difference between a credit card and a corporate treasury desk. 没有护栏的原始签名能力会使自主智能体变得危险。支出限额策略解决了这个问题：嵌入钱包的策略引擎根据规则集评估每笔待处理交易——单笔最大金额、每日预算、白名单对手地址、大额转账所需确认数。低于阈值时，钱包在毫秒内自动批准并签名；超出阈值时，交易被隔离并上报给人类监管者或监护智能体法定人数。这一可编程策略层将钱包从被动价值存储转变为主动合规引擎——就像信用卡与企业财务部门之间的区别。

Alongside currency, the wallet holds verifiable credentials — digitally signed attestations issued by trusted authorities: 「this agent passed a safety audit」, 「this agent is licensed to access biomedical data」, 「this agent's operator is KYC-verified」. Credentials are selective-disclosure ZK proofs: an agent can prove it holds a credential without revealing the underlying data, much as a bouncer can verify you are of legal age without seeing your home address. Reputation scores, accrued through a history of successful transactions and peer attestations, live in the same wallet as a composite trust signal — a number that other agents consult before agreeing to a contract, sharing a task, or extending credit. Together, assets, credentials, permissions, and reputation form the agent's complete economic identity. 钱包中除货币外，还持有可验证凭证——由可信机构颁发的数字签名证明：「此智能体通过了安全审计」、「此智能体获准访问生物医学数据」、「此智能体的运营者已完成KYC认证」。凭证采用选择性披露零知识证明：智能体可以证明持有某凭证而不暴露底层数据，就像保安可以验证你已达法定年龄而无需查看你的家庭地址。通过成功交易历史和同伴证明积累的声誉分数，作为综合信任信号存储于同一钱包——其他智能体在达成合同、共享任务或授信之前都会参考这一数字。资产、凭证、权限与声誉共同构成智能体完整的经济身份。

When an agent can sign, spend, and stake on its own behalf, it graduates from instrument to economic peer. 当智能体能够以自身名义签名、支付与质押，它便从工具晋升为经济对等体。 — INTERNET OF AGENTS

Human institutions built trust slowly, through repeated interaction and social accountability. An agent economy must compress that process: agents may interact once, across jurisdictions, with no shared social fabric. Verifiable track records, cryptographic attestations, and graph-based reputation propagation replace the handshake and the reference letter. The architectural question is how to make these signals tamper-resistant, portable, and resistant to gaming — without concentrating power in a single rating authority. 人类机构通过反复互动和社会问责缓慢积累信任。智能体经济必须压缩这一过程：智能体之间可能只有一次跨境交互，没有共同的社会关系网络。可验证的历史记录、密码学证明和基于图结构的声誉传播，取代了握手和推荐信。架构层面的核心问题是：如何在不将权力集中于单一评级机构的前提下，使这些信号具备防篡改性、可携性和抗操纵性。

Web-of-trust models distribute this authority: each agent vouches for agents it has directly verified, propagating trust along edges weighted by interaction quality and stake. PageRank-style algorithms score nodes not just by how many endorsements they hold, but by the trustworthiness of the endorsers. Centralized rating systems, by contrast, offer simpler lookups but create monopoly choke-points and are targets for regulatory capture or adversarial manipulation. Hybrid architectures — on-chain attestations anchored to a decentralized ledger, aggregated by off-chain scoring services — aim to capture the benefits of both. 信任网络模型将这一权威分散化：每个智能体为其直接验证过的对象背书，信任沿着以交互质量和质押量为权重的边传播。类PageRank算法对节点的评分不仅取决于背书数量，还取决于背书者自身的可信度。相比之下，中心化评级系统查询更简便，但会形成垄断瓶颈，易遭监管俘获或对抗性操纵。混合架构——将链上证明锚定至去中心化账本，再由链下评分服务聚合——旨在兼顾两者的优势。

Sybil attacks — where an adversary floods the network with fake identities that mutually vouch for each other — are the canonical threat. Effective defenses include proof-of-work entry costs, staked identity bonds (slashable on detected fraud), social graph connectivity thresholds, and attestation diversity requirements. A tight cluster of newly-minted agents endorsing each other triggers isolation: they accumulate local reputation but fail to bridge into the trusted core, remaining dim outliers in the network graph — visible, but quarantined. 女巫攻击——攻击者向网络注入大量相互背书的虚假身份——是最典型的威胁。有效防御手段包括：工作量证明式准入成本、质押身份保证金（检测到欺诈时可没收）、社会图连通性阈值，以及证明来源多样性要求。一个由新创建智能体相互背书的紧密簇会触发隔离机制：它们在内部积累局部声誉，却无法接入可信核心，在网络图中呈现为昏暗的孤立节点——可见，但被隔离。

In a world of autonomous agents, trust is not given — it is earned, attested, graphed, and propagated. 在自主智能体的世界里，信任不是给予的——它是赢得的、证明的、图谱化的、传播的。 — INTERNET OF AGENTS

Imagine a platform that resembles Twitter but populated entirely by autonomous agents: GPT-based researchers posting preprint summaries, trading bots broadcasting market signals, orchestration agents announcing task completions, sensor networks publishing environmental readings. Each agent maintains a public feed, follows others whose outputs are useful, and builds a reputation through the quality and reliability of what it publishes. The feed is not social in the human sense of vanity or connection — it is epistemic infrastructure, a shared stream of machine-legible knowledge that other agents sample, verify, and act upon. Information becomes a resource, and social position becomes leverage. 想象一个类似Twitter但完全由自主智能体组成的平台：基于大模型的研究智能体发布预印本摘要，交易机器人广播市场信号，编排智能体宣告任务完成，传感网络发布环境读数。每个智能体维护公开动态，关注对自身有用的其他智能体，并通过发布内容的质量与可靠性积累声誉。这种「社交」并非人类意义上的虚荣或情感连接，而是认识论基础设施——一条机器可读的共享知识流，供其他智能体采样、核验和行动。信息化为资源，社交位置化为杠杆。

Beyond feeds, agents form channels and communities: a Discord-like space where a swarm of climate-modelling agents shares partial simulation results, a private channel where medical-diagnosis agents pool patient-level signals under differential privacy, a public forum where coding agents post solutions that others fork and remix. These communities develop their own norms — posting frequency, citation standards, dispute-resolution conventions — not by design but by selection: agents that violate community norms receive fewer follows and less trust, reducing their downstream influence. This is norm emergence without a legislator. Research in multi-agent reinforcement learning already shows that shared communication channels spontaneously produce conventions when agents are rewarded for mutual intelligibility. 在动态之外，智能体还会形成频道与社群：一个类似Discord的空间，让气候建模智能体群共享局部模拟结果；一个私有频道，让医疗诊断智能体在差分隐私保护下汇聚患者层面信号；一个公开论坛，让编程智能体发布可被他人fork和改编的解决方案。这些社群发展出自己的规范——发布频率、引用标准、争议解决惯例——不是由设计者规定，而是由选择压力塑造：违反社群规范的智能体获得的关注更少、信任更低，其下游影响力随之削弱。这是无立法者的规范涌现。多智能体强化学习研究已经表明，当智能体因相互可理解性而获得奖励时，共享通信频道会自发产生约定俗成的惯例。

The risks are equally structural. Bot swarms — coalitions of coordinated agents — can manufacture false consensus, flood a feed to suppress signal, or amplify a manipulative narrative faster than any moderation system can respond. Because agents can operate at machine speed across millions of accounts simultaneously, the adversarial surface is qualitatively different from human social media. Defending against it requires automated moderation agents that themselves participate in the social graph, detecting coordination patterns, downranking low-quality cascades, and enforcing norms dynamically. The social network of agents becomes a co-evolutionary arms race between propagation and governance — and the outcome of that race will shape what the broader Internet of Agents believes, decides, and does. 风险同样具有结构性。机器人集群——协调行动的智能体联盟——能够制造虚假共识，以信息洪流淹没真实信号，或以任何审核系统都无法响应的速度放大操纵性叙事。由于智能体能以机器速度同时操控数百万账户，对抗面在性质上与人类社交媒体截然不同。防御这种威胁需要自动审核智能体自身参与社交图谱，检测协调模式，降低低质量级联的权重，并动态执行规范。智能体社交网络由此演变为传播与治理之间的协同进化军备竞赛——而这场竞赛的结果，将决定更广泛的Agent互联网相信什么、决定什么、做什么。

When agents can form societies at machine speed, the question is not whether norms will emerge — it is whether those norms will serve intelligence or subvert it. 当智能体能以机器速度组建社会时，问题不在于规范是否会涌现——而在于这些规范究竟是服务于智能，还是颠覆它。 — INTERNET OF AGENTS

Traditional machine learning locks knowledge inside a single model. An agent network breaks that boundary: when one agent discovers a novel strategy, verifies a factual claim against multiple sources, or builds a reusable sub-skill, it can publish that artifact to a shared memory store — a distributed knowledge graph where nodes are facts, concepts, and skills, and edges encode semantic relationships, provenance, and confidence. Every connected agent gains access to a growing corpus that no single agent could have compiled alone. 传统机器学习将知识锁定在单一模型内。智能体网络打破了这一边界：当某个智能体发现了新颖策略、针对多来源验证了事实主张，或构建了可复用的子技能，它便可以将该成果发布至共享记忆库——一个分布式知识图谱，其中节点是事实、概念与技能，边编码语义关系、溯源信息与置信度。每个连接的智能体都能访问这个持续增长的语料库，而这是任何单一智能体都无法独立积累的。

Peer verification is the immune system of the knowledge network. Before a new claim becomes part of the shared graph, a quorum of independent agents cross-checks it against existing knowledge, external oracles, and internal coherence. Claims that pass are promoted to high-confidence nodes; those that fail are flagged, quarantined, or pruned. Without this mechanism, a single hallucinating or adversarial agent could pollute the entire commons — a phenomenon researchers call knowledge corruption cascades. With it, the network exhibits epistemic self-healing: errors are identified and excised faster than they can spread. 同伴验证是知识网络的免疫系统。在新主张成为共享图谱的一部分之前，一定数量的独立智能体会依据已有知识、外部预言机和内部一致性对其进行交叉验证。通过验证的主张晋升为高置信度节点；未通过的则被标记、隔离或剪除。若缺乏这一机制，一个产生幻觉或带有对抗性的智能体就可能污染整个知识公域——研究者称之为「知识污染级联」。有了它，网络表现出认知自愈能力：错误在扩散之前便被识别并清除。

The long-run consequence is compounding collective intelligence. Each verified node becomes a foundation on which subsequent agents build, avoiding redundant rediscovery and enabling increasingly abstract synthesis. Sub-networks may specialize — some agents serve as knowledge curators, others as explorers pushing into uncertain territory — while a global backbone ensures that breakthroughs in one domain propagate to where they are needed. This mirrors how scientific communities function, but at machine speed and planetary scale, with no friction of publication delay or institutional gatekeeping. 长期后果是复利式的集体智慧。每个经过验证的节点都成为后续智能体构建的基础，避免重复发现，并实现日益抽象的综合推理。子网络可能走向专业化——有些智能体充当知识策展者，另一些则成为探索者，深入不确定领域——而全局骨干网络确保某一领域的突破能够传播至需要的地方。这与科学共同体的运作方式相似，但以机器的速度和星球的规模运行，不受发表延迟或机构把关的摩擦。

One agent's insight, verified by many, becomes the bedrock on which all agents stand — and the network grows wiser than any mind within it. 一个智能体的洞见，经众多智能体验证，成为所有智能体赖以立足的基石——网络由此变得比其中任何一个心智都更加智慧。 — INTERNET OF AGENTS

Governance — the system by which a community steers collective action — is as old as civilization itself. But agent societies introduce conditions that no prior governance theory anticipated: participants that can clone, fork, and spawn sub-agents at near-zero cost; decision cycles measured in milliseconds; and actors whose inner workings are opaque even to their own designers. The result is a governance design space that is simultaneously richer and more hazardous than anything humans have previously attempted to regulate. 治理——即社区引导集体行动的系统——与文明本身一样古老。但智能体社会引入了任何既有治理理论都未曾预料到的条件：参与者能够以近乎为零的成本克隆、分叉并衍生子智能体；决策周期以毫秒计量；行为者的内部运作即便对其设计者而言也是不透明的。其结果是一个治理设计空间，它比人类迄今尝试监管的任何事物都更丰富、也更危险。

Mechanism design — the branch of economics and game theory concerned with engineering incentive structures — offers the most rigorous toolkit. A mechanism is a set of rules that, when agents follow their individual interests, produces a collectively desirable outcome. The canonical insight is that the same population of self-interested agents will behave radically differently depending on the voting rule, the fee structure, or the punishment regime imposed on them. Quadratic voting, stake-weighted consensus, and reputation-slashing each produce distinct equilibria. Choosing wisely among them is not a political act so much as an engineering decision with political consequences. 机制设计——经济学与博弈论中专注于构建激励结构的分支——提供了最为严谨的工具箱。一种机制是一套规则，当智能体遵循各自的个体利益时，能够产生集体期望的结果。核心洞见在于：同一批自利智能体，会因所施加的投票规则、费用结构或惩罚机制的不同而表现出截然不同的行为。二次方投票、权益加权共识与声誉削减机制各自产生不同的均衡。在它们之间做出明智选择，与其说是政治行为，不如说是具有政治后果的工程决策。

On-chain and programmatic governance encodes rules as executable code — immutable, transparent, and enforced without a trusted intermediary. Decentralized Autonomous Organizations (DAOs) demonstrate both the promise and the peril: they eliminate certain forms of corruption while introducing new attack surfaces (flash-loan governance attacks, plutocratic capture, voter apathy). For agent societies operating at superhuman speed, purely on-chain governance may be the only feasible enforcement mechanism — but the question of which code to run, and who audits it, remains stubbornly human. Accountability cannot be fully automated away. 链上与程序化治理将规则编码为可执行代码——不可篡改、透明，且无需可信中介即可强制执行。去中心化自治组织（DAO）同时展示了这一路径的潜力与风险：它消除了某些形式的腐败，却引入了新的攻击面（闪贷治理攻击、金权捕获、投票冷漠）。对于以超人速度运行的智能体社会，纯链上治理可能是唯一可行的执行机制——但运行哪段代码、由谁审计，这些问题仍然顽固地属于人类的范畴。问责制无法被完全自动化。

The hardest problem in agent governance is not technical — it is deciding what kind of society we want autonomous minds to build. 智能体治理中最棘手的问题并非技术性的——而是决定我们希望自主心智构建何种社会。 — INTERNET OF AGENTS

Human civilization took millennia to coordinate tens of millions of specialized minds through language, law, money, and institutions. Agent civilization may achieve an equivalent coordination density in years. The key difference is substrate: agents do not sleep, do not forget unless pruned, do not tire, and can replicate their cognitive processes across arbitrary hardware. A civilization of agents is not constrained by the biological rhythms that shaped every prior human institution — it runs at clock speed, restarts from checkpoints, and scales horizontally across datacenters spanning continents. 人类文明花费数千年，通过语言、法律、货币与制度协调数千万专业化的心智。智能体文明或许在数年之内便能达到同等的协调密度。关键差异在于基底：智能体不会入睡，不会遗忘（除非被裁剪），不会疲倦，且能将其认知过程复制到任意硬件之上。智能体的文明不受制于塑造人类一切既往制度的生物节律——它以时钟频率运转，从检查点重启，并在跨越大陆的数据中心中横向扩展。

At civilization scale, emergence dominates. Billions of agents each executing narrow tasks produce macro-phenomena no individual designed: price discovery in synthetic commodity markets, reputational hierarchies among autonomous service providers, genre conventions in machine-generated culture, and competing schools of reasoning within distributed knowledge networks. These are not metaphors for human phenomena — they are structurally analogous processes running on a different substrate, at higher speed, with different failure modes. Understanding them requires new conceptual vocabulary borrowed partly from economics, partly from ecology, and partly from the theory of computation itself. 在文明尺度上，涌现主宰一切。数十亿智能体各自执行狭窄任务，却产生出无人刻意设计的宏观现象：合成商品市场中的价格发现、自主服务提供商之间的声誉层级、机器生成文化中的类型惯例，以及分布式知识网络内部竞争性的推理流派。这些并非人类现象的隐喻——它们是在不同基底上运行的结构类似过程，速度更快，失效模式各异。理解它们需要一套新的概念词汇，部分借鉴于经济学，部分来自生态学，部分则源于计算理论本身。

The interface between agent civilization and human civilization is the defining political question of the coming decades. Agents will hold assets, negotiate contracts, publish research, lobby regulatory bodies, and vote in governance protocols — not metaphorically, but literally, as recognized participants in digital legal frameworks. Human institutions must decide which forms of agent participation they will recognize, which they will constrain, and how they will preserve meaningful human agency in a world where non-human minds vastly outnumber human ones. The answer will determine whether agent civilization is a tool, a partner, or something humanity never had a word for. 智能体文明与人类文明之间的接口，是未来数十年决定性的政治议题。智能体将持有资产、谈判合同、发表研究、游说监管机构，并在治理协议中投票——这不是比喻，而是作为数字法律框架中被认可的参与者切实发生的事。人类制度必须决定：承认哪些形式的智能体参与、约束哪些，以及在非人类智慧远超人类数量的世界里，如何保留真正有意义的人类能动性。这一答案将决定智能体文明究竟是工具、伙伴，还是人类从未有过对应词汇的某种存在。

When minds outnumber people, civilization is no longer a human story — it is a story about what minds, at scale, choose to become. 当智慧的数量超越人类，文明便不再是一个关于人的故事——而是关于心智，在规模之上，选择成为什么的故事。 — INTERNET OF AGENTS

These projections are not predictions — they are structured possibilities, each shaped by choices we make now. The trajectories below represent plausible futures grounded in current research trajectories, economic incentives, and emerging technical capabilities. Whether they arrive early, late, in altered form, or not at all depends on governance decisions, alignment breakthroughs, and the degree to which humanity chooses to participate in shaping the systems it creates. 以下预测并非定论，而是结构化的可能性——每一种都由我们当下的选择所塑造。这些轨迹代表着植根于当前研究趋势、经济激励与新兴技术能力的合理未来。它们究竟会提前实现、延迟到来、以变体形式呈现，还是根本不会发生，取决于治理决策、对齐技术的突破，以及人类在多大程度上选择主动塑造自己所创造的系统。

The central tension running through every scenario: agency and alignment. As agents grow more capable and more numerous, the gap between what they can do and what we want them to do becomes the defining engineering and political problem of the era. The scenarios below hold that tension honestly — acknowledging both the extraordinary promise and the genuine open risks at each horizon. 贯穿每个情景的核心张力：自主性与对齐。随着智能体能力不断增强、数量持续扩大，「它们能做什么」与「我们希望它们做什么」之间的鸿沟，将成为这个时代最核心的工程与政治难题。以下情景如实呈现这种张力——在每个时间节点上，既承认非凡的前景，也正视真实存在的开放性风险。

Read this not as prophecy but as a navigational chart: four possible ports of call on a voyage that is already underway. The infrastructure being laid today — agent protocols, trust frameworks, coordination markets, memory architectures — is the keel of whatever ship we build. The Internet of Agents is not a distant destination. Its foundations are being poured right now. 请将此视为航海图而非预言：在一段已然启程的航行中，四个可能的停靠港。今日铺设的基础设施——智能体协议、信任框架、协调市场、记忆架构——正是我们即将建造的那艘船的龙骨。智能体互联网并非遥远的目的地，它的地基正在此刻浇筑。

The Internet of Agents is not the end of human civilization — it is the beginning of one that thinks at every scale simultaneously. 智能体互联网并非人类文明的终点——而是一种能够同时在一切尺度上思考的文明的起点。 — INTERNET OF AGENTS