The race to build AI agents that can operate smartphones is accelerating. Instead of relying on APIs or app integrations, these agents interact with mobile apps the same way humans do like reading the screen, tapping buttons, typing text, and navigating multi-step workflows.
Two products taking very different approaches are ByteDance's Doubao AI Phone and Mobilerun, the mobile AI agent infrastructure developed by Droidrun.
While both aim to solve the same problem, they are designed for different audiences.
Doubao packages AI into a dedicated smartphone with a tightly integrated experience. Mobilerun provides the infrastructure that lets developers connect AI models to Android or iOS devices, whether those devices are physical phones or cloud phones.
We spent time testing both approaches across a series of real-world tasks to understand where each performs well and where each still struggles.
Two Different Answers to the Same Problem
Doubao and Mobilerun are both trying to answer a simple question: can an AI agent operate the apps on a phone the way a person does like reading the screen, tapping, typing, moving through multi-step tasks instead of relying on api endpoints which are not always available.
The difference lies in how they approach the problem:
Doubao solves it through the operating system. The AI is deeply integrated into a dedicated smartphone designed around an agent-first experience.
Mobilerun solves it through infrastructure. Rather than shipping a phone, it provides a platform that allows developers to connect AI models like Claude, Hermes, or OpenClaw to Android or iOS devices and automate mobile workflows.
Quick Feature Comparison
| Feature | Doubao AI Phone | Mobilerun |
|---|---|---|
| Product Type | AI-powered smartphone | Mobile AI agent infrastructure |
| Hardware Required | Dedicated Doubao phone | Existing Android, iPhone, or cloud phone |
| Cloud Phone Support | No | Yes |
| Agent Support | Doubao ecosystem | OpenAI, Anthropic, Gemini, Ollama, DeepSeek, Hermes, OpenClaw, and others |
| Target Audience | Consumers | Developers & businesses |
| Human-in-the-loop | Yes | Yes |
| Cross-application Workflow | Yes | Yes |
| Setup Complexity | Easy | Medium |
| Scalable to Multiple Devices | No | Yes |
| Hardware | Nubia M153 | Can operate most mobile devices |
| Background Task Execution | Yes | Limited to cloud phones |
How Doubao Works
The first thing you notice on the Doubao AI Phone is its dedicated AI button. Press it and the assistant appears immediately, ready to execute the task. It opens apps, navigates interfaces, types text, and moves through multiple screens autonomously. The experience feels much closer to delegating work to an assistant.
How Mobilerun Works
Instead of requiring dedicated hardware, it connects AI models to existing mobile devices. Users can either:
- connect their own Android or iPhone or Cloud phones for remote scalable automation, or
- launch cloud phones for remote automation. The AI interacts with the device by observing the interface and performing actions on behalf of the user. Unlike consumer-focused assistants,Mobilerun is a infrastructure layer for mobile agents.
Let’s compare both Doubao and Mobilerun for a few tasks:
Ordering an Iced Americano: Doubao vs Mobilerun
On Doubao: We asked Doubao to order an iced Americano. The assistant opened the food delivery app, searched for coffee, compared prices, selected the lowest-priced option, and attempted to complete the order. During the process, it paused appropriately when clarification was needed.
The workflow completed successfully, although the entire process took close to three minutes. Overall, the experience felt similar to delegating the task to a junior assistant that occasionally asks follow-up questions before proceeding.
On Mobilerun: We repeated the same task using Mobilerun. The agent opened the delivery application, searched for an iced Americano, compared available options, selected one, and added it to the cart. Rather than automatically completing the purchase, it paused and asked whether the order should be placed or left in the cart.
This human-in-the-loop workflow provides an additional confirmation step before performing actions with financial consequences.
Task 2: Comment on a post
On Doubao: We asked the agent to post a comment under Instagram post, Doubao repeatedly typed and deleted text, then gave up, and simply couldn't identify the comment button. It's a small, telling failure: the agent handles familiar surfaces well, but a control it can't confidently locate stops it.
On Mobilerun. We then repeated the same task using Mobilerun. During our testing, the agent successfully navigated to the post, entered the requested comment, and submitted it.
One factor that likely contributed to this result is Mobilerun's ability to combine a11y tree information with computer vision when interpreting the interface, allowing it to identify actionable UI elements more reliably in many situations.
Task 3: Chat with a friend on Wechat
On Doubao: For our final test, we asked Doubao to open WeChat and send a message to a friend. During testing, the workflow did not complete successfully.
Public reports have noted compatibility challenges between the Doubao AI Phone and certain major Chinese applications, including WeChat, although the exact causes have not been fully disclosed publicly. Some applications reportedly blocked Doubao from automating tasks (due to security issues)
On Mobilerun: We repeated the same task using Mobilerun. This time, the agent understood the context, opened WeChat, and chatted with the friend. Again for this task, Mobilerun agent successfully executed the task, while Doubao struggled.
What these tasks reveal
Doubao is a genuine milestone: the first agent-led smartphone that is integrated into a user's device.
Unlike earlier swings such as the Rabbit R1, which stumbled on even its promised basics, Doubao completes real end-to-end workflows like ordering food, playing podcasts, navigating apps entirely through the agent, on a device you can buy today. But it runs those tasks as a self-contained device experience, and when it can't parse a UI element or hits an app's defenses, it stops. It works just well enough to prove an agent-first phone is possible, and fails just often enough to show how far the idea has to go. It is still a testing ground.
Mobilerun treats the same problem as a layer rather than a device. Because it's model-agnostic (OpenAI, Anthropic, Gemini, Ollama, DeepSeek) and runs on both Android and iOS, on personal phones or cloud phones, it isn't tied to one handset or one vendor's reasoning. The human-in-the-loop chat model means you stay in control of consequential actions like placing an order, making a payment etc. And the a11y + vision mode + better architecture is what lets it push through the exact UI moments, like finding a "Send" button, where Doubao agent gives up.
Doubao is a product you hold, Mobilerun is a infrastructure layer you operate with. But on the same tasks, they draw the two roads to the same destination: the phone as something you delegate to rather than operate.
Which one to choose?
If you're looking for a consumer-friendly AI phone that works out of the box, Doubao offers one of the most compelling implementations currently available.
If you're building AI agents, automating mobile workflows, or integrating smartphones into larger AI systems, Mobilerun provides considerably more flexibility and infrastructure.
Top comments (0)