IPADS-SAI has open sourced MobiAgent, a customizable agent framework for mobile GUIs. It is not a single model repository. Instead, it puts models, executors, acceleration mechanisms, benchmarks, and mobile apps into one system, with the goal of letting agents complete cross-app, multi-step tasks in real phone environments.
From the project structure, MobiAgent mainly consists of three parts: the MobiMind agent model series, the AgentRR recording-and-replay acceleration framework, and the MobiFlow benchmark. The paper abstract also emphasizes that accuracy and efficiency in real mobile tasks remain the main bottlenecks for current mobile agents, and MobiAgent is designed around those two problems.
What Problem It Solves
Mobile GUI agents are more troublesome than web or desktop automation. They need to understand screenshots, identify controls, decide the next action, and then use ADB or a mobile runtime to tap, type, go back, and switch apps. Real tasks are often not one operation inside one app, but a continuous flow across search, shopping, social, payment, maps, and other apps.
MobiAgent systematizes these pieces:
- MobiMind handles task planning, decision-making, and interface localization.
- The runner connects to the phone, executes predefined tasks through ADB, and records traces.
- AgentRR reuses successful action sequences to reduce reasoning and operation cost for repeated tasks.
- MobiFlow evaluates task completion in real mobile scenarios.
- Data collection, annotation, and processing tools lower the cost of building mobile GUI task data.
This makes it more like a mobile-agent experimentation platform than a model project that can only run demos.
Recent Updates Worth Watching
The README shows that MobiAgent was open sourced in August 2025 and then continued to fill in models, runner, memory system, and on-device execution capability. From December 2025, the project supported pure on-device inference on phones and released a unified GUI agent runner that can be configured with MobiAgent, UI-TARS, AutoGLM, Qwen-VL, Gemini, and other models.
By March 2026, the project had also released the GUI-based mobile “claw” MobiClaw and the new MobiMind-1.5-4B model. This suggests that it is not just reproducing a paper, but continuing to push mobile execution, model capability, and operation tooling toward a more product-like direction.
Memory Is a Key Patch
MobiAgent supports user profile memory, experience memory, and action memory. User profile memory gives planning preference context; experience memory retrieves execution experience from similar tasks; action memory uses AgentRR to cache and reuse successful action sequences.
This matters because phone tasks are naturally repetitive. Users often search products in the same app, open fixed contacts, or fill information on particular pages. If the agent has to inspect the screen, plan, and tap from scratch every time, the cost is high and errors are likely. Memory can preserve part of the “learned flow”, making later tasks faster and more stable.
Memory also creates governance questions. User preferences, task history, app paths, and action traces may contain sensitive information. In real deployments, the system needs to define what enters memory, how long it is stored, how it can be deleted, and whether the model may reuse that context across tasks.
Who Should Follow It
If you only want a ready-made phone automation app, MobiAgent is still more of a research and engineering framework. It requires model services, mobile devices, ADB, dependencies, and task files, so a full run has a real setup cost.
But if you care about mobile GUI agents, on-device agents, multi-model runners, task-trace reuse, or agent evaluation, MobiAgent is worth tracking. It places models, execution, evaluation, and data pipelines together, which helps researchers and developers see the real bottlenecks of mobile agents more completely.
My Take
MobiAgent matters not because it publishes one more GUI agent, but because it pushes phone agents beyond the single ability of “look at a screenshot and tap a button” into a framework that can be trained, executed, evaluated, and accelerated.
Mobile is a scenario agents cannot easily avoid. Many personal tasks happen inside apps rather than standardized web pages or APIs. Whoever can reliably understand phone interfaces, execute cross-app tasks, reuse experience, and control privacy risks will be closer to a truly usable personal agent.
MobiAgent has not solved all of these problems yet, but it provides a fairly complete open source starting point. In the short term, it is suitable for mobile-agent research and experimentation; in the long term, frameworks like this may become an important connection layer among mobile operating systems, personal assistants, and automation tools.
Project link: IPADS-SAI/MobiAgent
Paper link: MobiAgent: A Systematic Framework for Customizable Mobile Agents