The agent's next stop is your pocket: why mobile may be the agent's best form

A developer argues that mobile devices, not browsers, are the ideal platform for AI agents, citing the shift from desktop-bound interaction to asynchronous, decision-focused mobile use. The developer released DeepSeek Phone Harness, an open-source tool that lets users dispatch agents and return for results, embodying the principle that users shouldn't have to babysit agents.

We moved agents from the terminal to the browser. Now it's time to move them into your pocket. This isn't "mobile adaptation" — it's a shift in interaction paradigm. DeepSeek Harness just went open source, and the community is excited — models, tools, agent loops, all pluggable. But amid the noise I noticed something: Everyone's discussing what agents can do. Almost nobody discusseswhere you use them. Let's talk about the second one — and I'll give you my own answer. First generation: open a terminal, sit down, watch. Second generation Web GUI : open a browser, sit down, watch. Essentially it's still "you must be seated at the computer." But an agent's whole value is that it does the work for you — and yet you have to babysit it. Is that reasonable? All these share one trait: you're not at the computer, but the agent is working. Who's the remote? The phone is the only answer. I used to think mobile = shrunk web. After actually building it, I found mobile's interaction logic is fundamentally different: 1. From "multiple windows" to "a single timeline" The web has panels side by side: conversation, tool output, approval dialogs. Mobile naturally has one vertical timeline — user message, agent reasoning, tool cards, approval cards, question cards, all linear. The agent's work becomes a replayable pipeline , which is clearer, not less. 2. From "clicking around" to "making decisions" On desktop you click everywhere. On mobile, thumb reach is limited, which forces interactions into "look once → decide" : allow this tool call or not? How to answer this question? — judgment calls, not browsing sessions. 3. From "accompanying" to "async" On desktop you sit with the agent. On mobile you dispatch → leave → come back for results. That pushes the agent side to do state management properly task queues, result retention, failure notifications — which is exactly where agent products are heading. Based on the above, I built DeepSeek Phone Harness https://github.com/2903077918-lgtm/DeepSeek-phone-harness : It's no revolutionary invention — just a plain answer: "when the agent works for you, you shouldn't have to be there." What do you think is the agent's next form? Comment below — this is a topic worth arguing about.