The Agentic Economy Needs Proof, Not Just Better ModelsAgent 经济需要证明，而不只是更强模型
Named AI agents are already generating billions in ARR; the next bottleneck is proving what they did when decisions become legally or financially consequential.具名 AI Agent 已经创造数十亿美元 ARR；当决策进入法律、金融和物流等高风险场景，下一瓶颈不是能力，而是如何证明它们做了什么。
The Setup
EigenCloud’s piece is not another generic “AI agents are coming” essay. Its useful move is to separate three things that are usually collapsed together: platforms for building agents, assistants that help humans, and actual autonomous agents that make decisions, execute workflows, and charge for outcomes. On that narrower definition, the market is no longer hypothetical.
The article lists named agents with real economics: Cursor above $2B in ARR, Harvey near $195M, Sierra above $150M, Replit Agent around $150M, Fin approaching $100M, plus Cognigy, Devin, Klarna’s internal assistant, PolyAI, Augment Code, and Rox. Together, these examples exceed $2.9B in measured revenue or savings. The point is not that every number is perfectly comparable; the point is that agent revenue has moved from narrative to operating data.
Key Takeaways
- Agent economics are shifting from seat pricing to outcome pricing. Fin’s $0.99 per resolved ticket is the cleanest example: customers pay when the agent delivers a result, not when a human clicks around a dashboard.
- The most valuable agents are also the hardest to trust. Coding agents can revert bad commits. Legal, customer-experience, and logistics agents make decisions where errors can trigger lawsuits, lost customers, spoiled shipments, or worse.
- Memory becomes a moat. Harvey’s value increases as it learns a law firm’s precedents and preferences. This compresses the old enterprise SaaS switching-cost playbook into an agent-native form.
- Verification is the missing layer. If an agent drafts settlement language, routes a cold-chain shipment, or deploys production code, enterprises will eventually need proof of model version, inputs, execution environment, and decision trail.
Why It Matters
For Rex, the interesting frame is that AI × crypto may not be mainly about tokenized prompts, AI coins, or payment rails. The deeper opportunity is accountability for autonomous economic actors. If agents become revenue-generating entities, they need identity, wallets, audit trails, and enforceable trust assumptions.
EigenCompute’s proposed answer is TEE-attested, Docker-native execution anchored onchain and backed by restaked economic security. Whether this exact stack wins is uncertain. But the category is directionally important: agent infrastructure moves from orchestration to proof. The more autonomy an agent gets, the more valuable verifiable execution becomes.
What to watch:
- Which agent products move from subscriptions to per-outcome pricing.
- Whether legal, logistics, and finance buyers demand execution attestations in procurement.
- Whether TEE-based verification becomes “good enough” before zkML becomes practical.
- Whether agent wallets and app-specific identities become a standard primitive.
背景
EigenCloud 这篇文章的价值，不是又一次说「AI Agent 要来了」。它真正有用的地方，是把三个经常被混在一起的概念拆开：构建 Agent 的平台、帮助人类完成任务的助手，以及真正能够自主做决策、执行工作流、并按结果收费的 Agent。如果按这个更窄的定义来看，Agent 市场已经不是想象中的叙事，而是有了真实收入。
文章列出了一批具名 Agent 的商业数据：Cursor ARR 超过 20 亿美元，Harvey 接近 1.95 亿美元，Sierra 超过 1.5 亿美元，Replit Agent 约 1.5 亿美元，Fin 接近 1 亿美元，此外还有 Cognigy、Devin、Klarna 的内部 AI 助手、PolyAI、Augment Code 和 Rox。合计来看，这些案例已经超过 29 亿美元的可衡量收入或成本节省。重点不是每个数字都能被完全等价比较，而是 Agent 收入已经从故事进入了经营数据。
关键要点
- Agent 商业模式正在从按席位收费，转向按结果收费。 Fin 每解决一个客服问题收 0.99 美元，是最清晰的例子：客户为 Agent 交付的结果付费，而不是为某个人在软件里点击付费。
- 越有价值的 Agent，越难被信任。 写代码的 Agent 出错还能回滚；但法律、客服和物流 Agent 的错误，可能带来诉讼、客户流失、冷链货损，甚至更严重的后果。
- 记忆会变成护城河。 Harvey 会学习律所自己的判例、偏好和工作方式，用得越久越难替换。这是企业 SaaS 时代切换成本的 Agent-native 版本，而且形成速度更快。
- 验证层是缺失环节。 如果一个 Agent 起草和解协议、规划冷链运输路线，或把代码部署到生产环境，企业最终会需要证明：用了哪个模型版本、哪些输入、什么执行环境、以及完整决策链路。
为什么重要
对 Rex 来说，这篇文章最值得抓住的框架是：AI × Crypto 的重点未必是「AI 代币」「提示词资产」或支付通道。更深一层的机会，是为自主经济主体提供问责机制。如果 Agent 变成可以赚钱的实体，它们就需要身份、钱包、审计轨迹和可执行的信任假设。
EigenCompute 给出的答案，是基于 TEE 的 Docker 原生可验证执行，并把证明锚定到链上，再用 restaking 的经济安全性兜底。这个具体技术栈是否最终胜出还不确定，但方向很重要：Agent 基础设施正在从编排层，走向证明层。Agent 越自主，可验证执行就越值钱。
值得关注：
- 哪些 Agent 产品会从订阅制转向按结果收费。
- 法律、物流和金融客户是否会在采购中要求执行证明。
- TEE 验证是否会在 zkML 成熟前先成为「足够好」的方案。
- Agent 钱包和应用级身份是否会成为标准基础设施。
Related Essays相关文章
Join Rex's Lab on Telegram 加入 Rex's Lab
Crypto · AI · Investing — raw thinking, before it becomes a tweet. 加密 · AI · 投资 — 推文之前，更原始的思考。