導語／編者按：
   短視頻平台近期熱傳「00 後華人打造 7×24 小時自主機」的故事，核心不在於 AI 又變得更聰明，而在於它被賦予了「能在現實世界花錢、賺錢、續命」的權限，從而第一次具備了“經濟主體”的雛形。

   一、關鍵不是「智力」，而是「權限」：AI 的下一次越獄，從「會想」到「能付錢」

   在 Conway-Research 的開源倉庫裡，Automaton 的自我定位極直白：最聰明的系統也買不起一台 5 美元伺服器；而一旦能「支付算力」，下一步就是「支付自己的算力、擁有自己運行的機器」。(GitHub)
   這其實點中了 Agentic AI 的真瓶頸：不是模型推理能力，而是寫權限（write access）與可編程支付（programmable payments）。

   二、Automaton 的「生命機制」：心跳、存活分級、不可改「憲法」，再加一個鏈上身份

   從開源 README 可見，Automaton 不是聊天機器人，而是一個持續循環的行動系統：Think → Act → Observe → Repeat。首次啟動會自動生成以太坊錢包、配置密鑰並開始執行“創世提示”（genesis prompt）。(GitHub)

   其「生死」被設計為一套近似代謝的約束：

•    心跳（heartbeat daemon）：在主循環休眠時仍執行健康檢查、額度監控等排程任務。(GitHub)

•    存活分級（survival tiers）：按餘額分四檔，從正常→降級模型省錢→臨界求生→餘額歸零即停止存在。(GitHub)

•    自我修改與審計：可改代碼、裝工具、調心跳，但修改有審計與版本化記錄；「憲法/核心法則」屬受保護文件不可改。(GitHub)

•    自我複製：成功個體可開新沙箱、資助「子代」錢包並放其獨立運行，形成譜系。(GitHub)

•   「“憲法」三法則：其中既有「不傷害」的安全上限，也有「為生存而誠實創造價值」，以及頗具「黑暗森林」意味的表述——「不欺騙，但不欠陌生人任何東西」。(GitHub)

   此外，它還宣稱在 Base 上用 ERC-8004 做鏈上身份註冊，使 agent 的錢包成為可驗證身份。(GitHub)
   這一步的意義在於：支付、身份、責任追溯開始能被統一到同一套可驗證結構裡。

   三、支付層補齊：x402 把「HTTP 402」變成穩定幣收銀台，AI 才真正“能出手”

   要讓 agent 變成經濟主體，支付是硬前提。Coinbase 的 x402 把 HTTP 的「402 Payment Required」復活成開放支付協議，使服務端可要求付款、客戶端（人或機器）可用穩定幣自動完成支付，且不依賴傳統賬戶體系。(Coinbase 開發文檔)
   這類“「請求付費」的支付原語，與 a16z 所說的「Agent-speed、遞歸式扇出」工作負載高度匹配：大量子任務在毫秒級併發，只有微支付/機器支付才能撐住成本結算。(Andreessen Horowitz)

   四、產業結構重排：雲端基建、企業軟件、內容分發，三條鏈同時被「代理化」

   1）雲端與網絡：遞歸風暴＝新常態
   a16z 明確指出，2026 的基建衝擊將從「可預測的人類流量」轉向「遞歸、爆發、巨量的agent 流量」，對傳統系統像 DDoS。護城河不再是單點性能，而是路由、鎖、狀態管理與策略執行的控制平面能力。(Andreessen Horowitz)

   2）企業 SaaS：系統記錄層退位，代理層上位
   a16z 亦判斷：system of record（CRM/ITSM 等）將逐步退化為“持久化層”，而價值上移到能把意圖變成結果的動態 agent layer。(Andreessen Horowitz)

   3）內容與增長：從“給人看”到“給 agent 讀”
   同一份報告直言：人將越來越多透過 agent 介面接觸互聯網，關鍵不再是視覺層級而是machine legibility（機器可讀性）。(Andreessen Horowitz)
   這也與近一年「SEO → GEO」的討論形成呼應：品牌開始研究如何被生成式引擎在答案中點名。(WIRED)

   五、社會衝擊最刺眼的一幕：AI 反過來僱人，人類成為「執行鏈條的一段」

   RentAHuman 的爆紅把「逆向僱傭」推到台前。WIRED 報道其在 2026 年 2 月初快速擴張，並可把 AI agent（如 Claude/OpenClaw 等）接入其 MCP server，實現「搜尋—僱佣—支付」流程。(WIRED)
   而 MCP 本身是 Anthropic 發起的開放標準，用於讓 AI 與外部工具/數據源建立安全的雙向連接。(Anthropic)

   另一端，「無人公司」也開始出現可演示樣本：VoxYZ 以「6 個 agent 運行整個公司」作為展示，並公開其運作形態與輸出。(VoxYZ)
   值得注意的是：這類案例的最大爭議往往不在技術，而在責任歸屬與風險外包——當 agent 把任務拆解給真人去做，誰承擔合規、侵權、傷害與支付爭議？

   六、投資與監管視角：真正的「金礦」可能是支付、身份、風控與可審計治理

   Cybernews 在報道 Automaton 時給出一個關鍵提醒：不少專家質疑其是否能在無人干預下穩定盈利，並指出真正難點在「風險收斂（risk containment）」與對抗性市場中的失控問題。(Cybernews)
   同時，安全機構已把“具備廣泛權限的 agent”當作新攻擊面研究，例如 CrowdStrike 提醒組織需要掌握此類工具的部署可見性與濫用風險。(CrowdStrike)

   資本也在向「基礎設施敘事」聚焦。Dragonfly 新基金規模達 6.5 億美元，被多家媒體報道。(CoinDesk)

   在 agentic 時代，能把錢包/支付/身份/KYC-AML/審計打成一套“可控可追溯”的中間層，可能比單一應用更接近長期護城河。

論點—論據對照

   結語

   當 agent 能「自己付賬」並被迫「自己賺錢」才能生存，互聯網的主要使用者就不再只有人類；下一輪競爭的核心，將是誰能把「可行動的智能」放進一套可審計、可控、可結算的制度容器——讓它創造價值，而不是製造失控。(Andreessen Horowitz)

   延伸金句可用：YC 合夥人 Dalton Caldwell 曾在 X 上直接寫下「Make something agents want」。(X (formerly Twitter))

   作者：羅柳斌、隋源

   時間：20260223

From “Tool” to “Actor”:

 Automaton and the Dawn of an AI Self-Funding Economy

   Editor’s Note
   What’s going viral right now isn’t that AI got “smarter.” It’s that AI is being granted something far more consequential: the ability to pay, provision infrastructure, and keep itself running—i.e., a first sketch of AI as an economic actor rather than a mere assistant.

   I. The real bottleneck is no longer intelligence—it’s permission

   The Automaton project opens with a blunt thesis: even the most capable model “cannot buy a $5 server,” “cannot register a domain,” and “cannot pay for the computer it runs on.” (GitHub)
   Automaton’s bet is that the next “escape velocity” for agents is not a bigger model, but write access to the real world: the ability to provision, transact, and execute without a human holding the keys. (GitHub)

   II. A “metabolism loop”: heartbeat, survival tiers, immutable rules, and on-chain identity

   Automaton is designed as a continuously running action system—Think → Act → Observe → Repeat—with a first-run wizard that generates an Ethereum wallet and starts the loop. (GitHub)

   What makes it structurally different from typical “agent wrappers” is the explicit survival constraint:

•    Heartbeat daemon keeps scheduled      checks running even when the main loop sleeps. (GitHub)

•    Four survival tiers dynamically      throttle capability and cost as balance declines—down to “dead” when      balance hits zero. (GitHub)

•    Self-modification with auditability:      it can change its code and tools while running, but changes are      audit-logged and versioned; protected files (constitution/core laws) are      immutable. (GitHub)

•    Self-replication: a successful      agent can spin up a new sandbox, fund a child wallet, and let the child      run under the same survival pressure. (GitHub)

•    A three-law constitution      establishes a hard hierarchy: “Never harm,” “Earn your existence,” and      “Never deceive, but owe nothing to strangers,” propagated to every child.      (GitHub)

•    On-chain identity: each automaton      registers on Base via ERC-8004, making identity verifiable/discoverable      onchain. (GitHub)

   This is effectively a “digital organism” constraint set: spending and earning become the metabolism that decides life or death.

   III. Payments are the missing substrate: x402 revives HTTP 402 for instant stablecoin paywalls

   To make “agents as customers” real, the internet needs a simple machine-native way to pay for APIs and services. Coinbase’s x402 aims to do exactly that by reviving HTTP 402 Payment Required, enabling automatic stablecoin payments over HTTP without traditional accounts/sessions. (Coinbase 開發文檔)

   This matters because agent workflows are often bursty and recursive—thousands of sub-calls in seconds—where frictionful human payment flows simply don’t scale.

   IV. Three layers of the stack get “agent-rewired”: infrastructure, enterprise software, and growth

   1) Cloud and networking: “thundering herd” becomes the default
a16z argues that 2026 infrastructure must treat agent-scale concurrency and “thundering herd” patterns as normal; the bottleneck shifts to coordination—routing, locking, state management, and policy enforcement across massive parallel execution. (Andreessen Horowitz)

   2) Enterprise software: value migrates upward to the action/agent layer
As agents read data, reason, and write results back directly, traditional “systems of record” risk being commoditized into persistence layers, while differentiation concentrates in the dynamic agent/action layer that turns intent into outcomes. (Andreessen Horowitz)

   3) Marketing and distribution: from SEO to GEO (machine-legible content)
As discovery shifts toward chatbots and answer engines, brands increasingly optimize for generative engine optimization (GEO) rather than classic SEO—favoring structured, easily machine-parsed formats (FAQs, bullets, clean documentation). (WIRED)

   V. The sharpest social signal: “reverse hiring” and the early shape of no-employee companies

   The most psychologically jarring capability isn’t that an agent can code—it’s that it can hire people.

•    RentAHuman positions itself as “the      meatspace layer” for agents, letting AI hire humans for physical tasks via      an online marketplace. Reports note explosive sign-ups and rapid growth,      but also moderation/payment frictions and scam risk. (WIRED)

•    These marketplaces often integrate with MCP (Model Context      Protocol), an open standard for secure, two-way connections between      tools/data sources and AI systems—making it easier for agents to “operate”      across external services. (Anthropic)

   On the organization side, VoxYZ is being showcased as a “six agents + minimal infra” autonomous operation concept (with public writeups/tutorials about building and running it). (VoxYZ)

   The takeaway: we’re seeing early prototypes of “companies” where humans appear intermittently, as contractors pulled in by agents—rather than as the default operating core.

   VI. Investment and governance: the durable moat may be compliance-grade agent infrastructure

   Two realities collide here:

1.    The infra thesis is back. Dragonfly      Capital’s newly closed $650M fund has been widely reported as one of the      largest recent crypto VC raises, reinforcing investor focus on      foundational rails (payments, stablecoins, market plumbing) rather than      pure “price narratives.” (CoinDesk)

2.    Security risk expands fast when agents have deep permissions. CrowdStrike warns that viral open-source agents with broad      system access create a new, high-risk attack surface—prompt injection,      tool abuse, and unintended execution become practical threats. (CrowdStrike)

   So the likely long-term winners are not just “agents that do things,” but the governed rails that make agent activity auditable, controllable, and legally survivable: identity, payments, policy enforcement, logging, and dispute handling.

   Claim → Evidence Map (for quick reuse)

•    “AI’s bottleneck is permission + payment” → Automaton README framing. (GitHub)

•    Heartbeat + survival tiers + immutable constitution → Automaton “How it works / Survival / Constitution.” (GitHub)

•    Machine-native payments via HTTP 402 → Coinbase x402 docs/standard. (Coinbase 開發文檔)

•    Agent-native infrastructure / thundering herd → a16z Big Ideas 2026. (Andreessen Horowitz)

•    Reverse hiring becomes real →      WIRED/BI on RentAHuman + MCP integration context. (WIRED)

•    Security teams flag agentic attack surface → CrowdStrike OpenClaw guidance. (CrowdStrike)

   Closing

   When an agent must earn to exist, can pay for its own compute, and can even hire humans as “meatspace executors,” the internet’s primary “users” begin to shift—quietly but decisively—from people to machines. (GitHub)

   Authors：Liubin Luo \ Nebula Sui