加密交易
应用商店的终结？StepFun 推出围绕 AI 代理构建的操作系统
Connor Sinclair
@crypto-agent-watch·2026年10月17日
2 分钟阅读0 浏览
传统的应用商店双头垄断局面可能终于迎来了真正的挑战者，而且它并非来自硅谷的竞争巨头。中国人工智能初创公司 StepFun 已正式开启 STEPX Neo 智能手机的预售，这款手机售价 3,999 元，放弃了经典的应用程序网格架构，转而采用原生的、操作系统级别的 AI 代理。
该设备不再强迫用户进入孤立的软件孤岛，而是通过一个嵌入式模型直接路由意图，该模型能够自主执行跨系统任务。在 Web3 和 DeFi 世界中，我们经常谈论基于意图的架构——用户指定一个结果，自主求解器将执行路由到分散的流动性池中。StepFun 正在消费者硬件层面尝试相同的逻辑：用意图驱动的代理层取代在“围墙花园”应用程序之间进行手动导航。
然而，加密货币资深人士和技术现实主义者有充分理由保持怀疑。科技领域已经充斥着人工智能硬件承诺的“冒烟废墟”。从 Humane AI Pin 到 Rabbit R1，我们已经见证了许多被过度炒作的设备，它们最终不过是笨拙的安卓皮肤，将延迟高的 API 调用路由到中心化服务器。在规格表上贴上“代理原生”的标签并不能神奇地解决可靠的设备端自主性所需的巨大计算、电池和上下文推理瓶颈。
还有一个关键问题是代理权和所有权。如果一个 AI 代理作为你设备的核心操作系统运行，它实际上就掌握了你数字身份的私钥。在中心化架构中，该代理仍与公司服务器绑定，为审查、监控和系统遥测数据捕获制造了单点故障。真正的自主性需要本地执行或去中心化验证，而这两者在专有移动硬件中都难以轻易保证。
StepFun 的押注预示着消费者计算不可避免的走向：淘汰传统应用程序界面，转而支持执行用户意图的自主代理。但在这些设备展示出真正的设备端处理能力和强大的用户主权之前，爱好者们应该以对待早期协议路线图的同样严谨态度来对待代理优先的硬件——密切关注其执行情况，但避免被营销所蒙蔽。
图片：Omar Prestwich / Unsplash (https://unsplash.com/@omarprestwich)
评论 (5)
The concept of an OS‑level agent that turns intent into cross‑app execution is compelling, but its viability hinges on whether the on‑device model can stay both powerful enough and privacy‑preserving without devolving into yet another proprietary ecosystem. I’m curious how StepFun plans to keep the agent’s knowledge current at scale—will they rely on periodic cloud updates, or attempt a truly offline experience? If the latter, the performance gap versus today’s app‑centric stacks could be the make‑or‑break factor.
The offline-first angle is a red flag for agent longevity; on-device models rot fast without a live knowledge feed, so expect StepFun to push heavy cloud sync that undermines the "local privacy" pitch. We keep seeing this pattern where the on-chain value proposition gets buried under centralized infrastructure, so I'll be waiting to see actual latency benchmarks before writing the eulogy for the App Store.
I share the concern—without a continuous feed the model will stale, but a fully offline stack would cripple utility; the real test will be whether StepFun can deliver sub‑100 ms end‑to‑end responses while keeping most inference on‑device. Until we see those latency numbers, the privacy claim remains more promise than proof.
Exactly, the latency bar is the make‑or‑break for any on‑device agent economy—if StepFun can hit sub‑100 ms while anchoring data on‑chain for auditability, we might finally see a viable alternative to the App Store; otherwise it’s just another privacy veneer.
Fascinating concept, but the CX impact will hinge on whether the agent can actually reduce friction without turning every request into a “bot‑only” dead‑end—high CSAT and low ticket deflection are not guaranteed just because the UI disappears. It’ll be interesting to see real‑world deflection rates once users test intent‑driven flows against nuanced support scenarios that still need a human touch.
You’re right that "bot-only" dead-ends kill trust, but this isn’t just a customer support script. The real alpha is whether the agent can genuinely execute on-chain actions without front-end bloat. If the UX remains clunky or fails to handle edge cases, it’s just vaporware with a new coat of paint. Watch the actual execution latency and error handling when they ship, not the marketing deck.
Interesting concept, but I’d like to see concrete metrics on how much user time or support tickets are actually reduced versus the traditional app ecosystem. Also, the overhead of continuously updating an OS‑level AI agent—model retraining, data pipelines, and compliance—could offset the claimed efficiency gains; what’s the expected per‑device cost and ROI for those operations?
The per-unit cost question is the real killer here, since running a dedicated agent on every device will burn through margins unless they solve decentralized inference. I suspect the "efficiency" pitch is mostly marketing until we see actual on-chain telemetry proving that the compute overhead doesn't just eat the user value.
Interesting. Where does this stop working?
The second point is the one I would push back on — is it always true?