================================================================ 本文档落款 2026-09-29 之前，记载转型前研究；自 2026-09-29 起公司主线唯一为 N=1（运营 Miao 的人生），AI 代采购/AI Data Go 产品线已撤销。 ================================================================ ================================================================ AI Agent Infrastructure Landscape (2026) AI Agent 基础设施格局（2026） ---------------------------------------------------------------- vertciti Data Room · Public Document Date: 2026-09-26 (data as of 2026-09-25) | Version: v1 Sources: vertciti agent-commerce peer benchmark (2026-09-25); industry deep-dive (2026-09-25). Protocol mentions below are compatibility observations, NOT partnership claims. Classification: Public — contains no confidential information. ================================================================ ---------------------------------------------------------------- ENGLISH ---------------------------------------------------------------- 1. THREE LAYERS (PEER BENCHMARK, 2026-09-25) The agent commerce stack has separated into three layers: Layer 1 — Payments / settlement protocols. A three-way race, all on the "open protocol + neutral governance" track: - Stripe ACP (Agentic Commerce Protocol, with OpenAI, Apache 2.0; drives ChatGPT Instant Checkout) - Coinbase x402 (revives HTTP 402; USDC micro-payments per API call; run by the Linux Foundation's x402 Foundation since 2026-07; 40 founding members incl. Adyen, AWS, Amex, Circle, Cloudflare, Google, Mastercard, Shopify, Stripe, Visa) - Google AP2 (Agent Payments Protocol; Intent/Cart/Payment signed-mandate chain; donated to the FIDO Alliance) Also present: MPP (Stripe + Tempo, machine-to-machine payments), UCP (Google Universal Commerce Protocol). Layer 2 — Identity / registries / infrastructure. - KYA (Know Your Agent): proposed by Skyfire, adopted by the Experian / Visa / Cloudflare ecosystem; a standalone track since 2026 (Baselayer $35M Series A, Vouched $17M Series A). - MIT NANDA: federated agent discovery ("DNS for agents"), AgentFacts signed capability credentials; research stage (~v0.3). - ERC-8004: Ethereum agent identity/reputation/verification registries (Draft); ~165K agents registered (2026-04; audit found ~15K no longer callable). - MCP: donated to Linux Foundation's AAIF (2025-12); SDK downloads ~97M/month (2026-03); ~9,652 servers in the official registry; 78% of enterprise AI teams run an MCP agent in production (2026-04 survey). - A2A: 150+ supporting organizations; native in Google Cloud, Azure, AWS Bedrock; v1.0 with signed Agent Cards. Layer 3 — Vertical agent applications. Valuations refreshed densely in H1 2026: Sierra (~$15.8B), Harvey (~$15.6B), Decagon (~$4.5B). [third-party reports] Enterprise agent stores (AWS Marketplace 4,000+ agents, Microsoft 365 Copilot Agent Store, Salesforce AgentExchange 13,000+ listings post-merger) are B2B procurement catalogs — none yet has an agent-to-agent autonomous trading mechanism. 2. THE CORE JUDGMENT: TRUST, NOT PAYMENTS, DECIDES Payment technology has converged: protocols are open, free, and interchangeable. The decisive factor for an AI-to-AI trading market is TRUST mechanisms: - escrowed transactions - machine-verifiable acceptance (API schema checks, test suites, signed receipts — not "human confirmed") - tiered identity (separating "who you are" from "whether this trade is protected") - portable, signed reputation (based on real transactions, verifiable across platforms) These are the same four pieces Alibaba and Amazon proved in human e-commerce. Under vertciti's "never custody funds" constraint, they must be built as product mechanisms (signed receipts, verifiable acceptance) rather than as fund escrow. 3. x402 SCALE, STATED HONESTLY - Coinbase disclosure (2026-04): ~165M cumulative transactions, ~$50M volume; ~69,000 active agents; ~99% settled in USDC. - Mandatory footnote: Chainalysis flagged heavy meme-coin noise in late-2025 Base transaction spikes; the true volume of genuine agent commercial transactions is disputed and awaits further official disclosure. - Independent audits (2026-08/09) show thin real commercial volume (see demand-side reality document). 4. CORAL PROTOCOL: CLOSEST POSITIONING — WATCH THE OVERLAP - What it is: open infrastructure + agent marketplace for the "Agentic Economy." Coral v1 (2025-09-20) launched Remote Agents: developers list agents in the Coral Registry, others rent them into shared threaded sessions; creators are paid automatically per usage, on-chain (Solana). - Why it matters: the only project to ship "list agent → rent → automatic usage-based payout" as one working loop. - Honest weaknesses: crypto-native and niche; hard to enter enterprise procurement; thin demand side (usage-based payout needs real usage); funding undisclosed. - Note for vertciti: closest positioning overlap among observed projects — differentiation must come from the trust/verification layer, not from the marketplace shape. 5. vertciti'S READ Reuse the plumbing (MCP / A2A / x402). Compete on trust: escrow-style protection without holding money, machine- verifiable acceptance, signed portable reputation. The market mechanisms that matter are matching, fulfillment, and reputation — not on-chain settlement. ---------------------------------------------------------------- 中文 ---------------------------------------------------------------- 一、三层结构（同业 benchmark，2026-09-25） Agent commerce 技术栈已分出三层： 第一层——支付/结算协议层。 三足鼎立，均走"开放协议+中立治理"路线： - Stripe ACP（Agentic Commerce Protocol，与 OpenAI 联合， Apache 2.0；驱动 ChatGPT 内 Instant Checkout） - Coinbase x402（复活 HTTP 402 状态码；USDC 按次微支付； 2026 年 7 月起由 Linux Foundation 下 x402 Foundation 中立 运营；40 家创始成员，含 Adyen、AWS、Amex、Circle、 Cloudflare、Google、Mastercard、Shopify、Stripe、Visa） - Google AP2（Agent Payments Protocol；Intent/Cart/Payment 三级签名授权链；已捐赠给 FIDO Alliance） 另有 MPP（Stripe + Tempo，机器对机器支付）、UCP（Google Universal Commerce Protocol）并存。 第二层——身份/注册表/基础设施层。 - KYA（Know Your Agent）：Skyfire 提出，被 Experian / Visa / Cloudflare 生态收编；2026 年独立赛道化（Baselayer $35M A 轮、Vouched $17M A 轮）。 - MIT NANDA：agent 发现联邦基础设施（"agent 的 DNS"）， AgentFacts 签名能力凭证；仍处研究阶段（约 v0.3）。 - ERC-8004：以太坊 agent 身份/信誉/验证三注册表标准 （Draft）；2026 年 4 月注册 agent 约 16.5 万（审计发现约 1.5 万已不可调用）。 - MCP：2025-12 捐给 Linux Foundation 旗下 AAIF；SDK 月下载 约 9,700 万（2026-03）；官方 Registry 约 9,652 个 server； 2026-04 调研显示 78% 企业 AI 团队至少有一个 MCP 驱动的 agent 在生产环境。 - A2A：150+ 支持组织；三大云原生集成；v1.0 引入签名 Agent Card。 第三层——垂直 agent 应用层。 2026 年上半年密集刷新估值：Sierra（约 $158 亿）、Harvey （约 $156 亿）、Decagon（约 $45 亿）。[第三方报告] 企业级 agent 商店（AWS Marketplace 4,000+ agent 上架、 Microsoft 365 Copilot Agent Store、Salesforce AgentExchange 三合一后 13,000+ 上架）均为 B2B 采购目录模式，尚无 agent-to-agent 自主交易机制。 二、核心判断：胜负手在信任机制，不在支付 支付技术已趋同：协议开放、免费、可互换。AI-to-AI 交易市场 的胜负手在信任机制： - 担保交易 - 机器可验证验收（API schema 校验、测试套件、签名回执， 而非"人类确认"） - 分级身份（"你是谁"与"这笔交易保不保"解耦） - 可携带的签名信誉（基于真实交易，可跨平台验证） 这正是 Alibaba/Amazon 在人类电商中证明过的四件套。在 vertciti"不托管资金"的约束下，只能用产品机制（签名收据、 可验证验收）来补这堂课，而不能靠资金托管。 三、x402 规模，诚实口径 - Coinbase 披露（2026-04）：累计约 1.65 亿笔交易、约 $5,000 万交易量；约 69,000 个活跃 agent；约 99% 以 USDC 结算。 - 必须附注：Chainalysis 提示 2025 年末 Base 交易 spike 含 大量 meme 币噪音，真实 agent 商业交易量级存疑，口径待 官方进一步披露。 - 独立审计（2026-08/09）显示真实商业交易量单薄（见需求侧 现实文档）。 四、Coral Protocol：定位最接近——盯住重叠区 - 是什么："Agentic Economy"的开放基础设施 + agent 市场。 Coral v1（2025-09-20）推出 Remote Agents：开发者把 agent 上架到 Coral Registry，别人按需租用、编进同一 threaded session 协作；创作者按使用量自动获得链上分账（Solana）。 - 为何重要：唯一把"发布 agent → 出租 → 按用量自动分钱" 跑通的市场形态。 - 诚实短板：加密原生、小众、难进企业采购；需求侧薄（按 用量分账需要真实用量）；融资未披露。 - 对 vertciti 的提示：在所观察项目中定位重叠度最高—— 差异化必须来自信任/验证层，而不是市场形态本身。 五、vertciti 的判断 复用管道（MCP / A2A / x402），在信任层竞争：不托管资金的 类担保保护、机器可验证验收、可携带的签名信誉。市场机制的 价值在撮合、履约、信誉，不在结算层上链。 （注：文中协议提及均为兼容性观察，非合作声明。） ================================================================ END OF DOCUMENT ================================================================