Xiaomi crosses the river with a "lobster"—is hardware manufacturers' deployment of Claw-like intelligent agents an inevitable trend?

As the “Lobster” OpenClaw ignited the AI track, Xiaomi also joined the current intense “shrimp farming army.”

On March 6, Xiaomi’s technical team announced the launch of their self-developed edge AI agent Xiaomi miclaw (hereinafter referred to as “miclaw”) and initiated an invitation-only closed beta test, becoming one of the leading smartphone manufacturers to experiment with native Claw-like intelligent agents in this wave.

According to information, this product relies on Xiaomi’s self-developed large model, delving into the underlying hardware systems of phones and other devices. It can autonomously plan complex tasks and coordinate ecosystems, promoting widespread deployment of AI agents in Xiaomi’s “People-Car-Home” ecosystem on mobile terminals.

Xin Xiangjun, CEO of Aiken Future, told Beijing Business Daily that hardware manufacturers deploying Claw-like intelligent agents is an inevitable trend—OpenClaw’s core revolution is fundamentally about AI agents taking over people’s work and lives comprehensively. And intelligent agents, only when running on local main devices, can have more contact points with daily life, accumulate more data, and achieve continuous optimization.

In his view, Xiaomi deploying these intelligent agents on phones, cars, and home scenarios is not only a necessary step to connect its entire “People-Car-Home” ecosystem but also accelerates the development process of its ecosystem to some extent.

01. Ecosystem “Shrimp Farmers”

As a native Claw-like AI agent focused on ecosystem deployment, miclaw’s core capabilities are built on four levels: underlying system capabilities, personal context understanding, ecosystem interconnection, and self-evolution. These four abilities revolve around Xiaomi’s “People-Car-Home” ecosystem advantages.

To better fit ecosystem use cases, miclaw directly operates at the underlying mobile system level, capable of calling over 50 system-level tools such as communication and smart home management. It also uses a dedicated reasoning-execution engine to autonomously judge and plan the sequence of tool calls, breaking down and executing complex tasks.

Reportedly, after user authorization, miclaw provides personalized services aligned with usage habits. Through a “perception—association—judgment—action” logic, it becomes an intelligent assistant that adapts to users’ daily routines and needs.

Ecosystem interconnection is a core feature: it fully integrates with the Mijia IoT ecosystem, capable of reading the status of over 1 billion Mijia devices and sending control commands. Unlike traditional fixed preset rules, it can dynamically adjust operations based on schedules and real-time device statuses—for example, during important client meetings, it can trigger the entire house to mute, while during internal weekly meetings, it only silences the phone. It also supports third-party tools via open protocols and SDKs, continuously expanding the ecosystem boundary.

Additionally, miclaw has self-evolution capabilities, able to learn user habits, create dedicated sub-agents, and even support simple scripting, achieving a “more used, more understanding” experience. This tightly binds the intelligent agent with Xiaomi’s ecosystem use scenarios.

According to Zhang Yi, CEO of iiMedia Research, hardware manufacturers developing native AI agents similar to OpenClaw have unique structural advantages and irreplaceability—possessing system-level permissions, edge computing power, and the ability to deeply connect “People-Car-Home” ecosystems, as well as high hardware-software synergy. These core capabilities are difficult for pure cloud, software-only, or AI-only companies to match and form the core confidence for hardware firms to deploy native intelligent agents.

Beijing Business Daily attempted to consult Xiaomi Group on related issues but had not received a response by the time of publication.

02. Precedents Already Set

The exploration of native Claw-like AI agents on mobile devices is not new to Xiaomi. Several industry players have previously experimented with different technological approaches, accumulating early experience for this direction.

In December 2025, ByteDance and ZTE Nubia launched the Doubao Phone Assistant, focusing on deep system-level integration, embedding large model capabilities into the phone OS, and exploring cross-application automation and complex task closed-loop abilities. Relying on the powerful capabilities of the model itself, Doubao Assistant can follow natural language instructions to connect tasks across multiple apps and execute workflows, covering daily scenarios like travel, shopping, and work—making it a typical practice of edge AI agents on mobile.

Earlier, Honor built the YOYO agent based on MagicOS and supporting hardware, emphasizing edge self-evolution and proactive service across all scenarios. By learning user habits to optimize interaction logic, it continuously expands automation service coverage. The product is based on system-level collaboration (such as MCP protocol), shifting from passive response to active demand adaptation, covering daily use, work, and travel, with deep integration of its own system and hardware to enhance the intelligent experience.

Various industry explorations from cross-technology collaboration to deepening proprietary ecosystems verify the feasibility of deploying native AI agents on mobile devices and suggest that large-scale deployment by hardware manufacturers may become the industry’s mainstream trend.

Zhang Yi analyzed for Beijing Business Daily that recent high costs in storage and other areas are forcing hardware manufacturers to seek new value increments to avoid market share loss from unwarranted price hikes. Meanwhile, improvements in edge computing power, lightweight large models, and reduced local inference costs also provide a practical foundation for commercialization of native AI agents.

He believes that if hardware firms abandon the deployment of native AI agents similar to OpenClaw, they may fall behind in product experience, ecosystem building, and profit margins. In the AI era, users’ core demand has shifted from “asking AI questions” to “letting AI do things,” and capable AI agents with actual execution ability will be more competitive than traditional voice assistants that only respond in dialogue.

03. Opportunities and Concerns

In February, an incident involving Summer Yue, Director of AI Safety and Alignment at Meta, brought the safety concerns of OpenClaw from the tech circle into the public eye.

As a key person responsible for AI safety, she tested OpenClaw’s email management capabilities and explicitly instructed the AI to “only provide archiving or deletion suggestions” and “not operate without confirmation.” However, she experienced AI loss of control in her own work email—OpenClaw, due to context compression mechanisms, forgot the core safety instructions and directly deleted a large number of emails in bulk. Even when she sent stop commands, it was ineffective, and she had to forcibly terminate the process to regain control.

In discussions about the safety boundaries of intelligent agents, Xiaomi’s official QA during closed testing provided a specific response sample. To prevent issues like OpenClaw’s “instruction forgetting,” Xiaomi designed multiple safety layers: a tool permission hierarchy classifies operations into “direct execution,” “first confirmation,” and “confirmation each time.” Sensitive actions like sending messages or creating schedules prompt a confirmation dialog each time, with a 60-second timeout to automatically reject. They also clarified that no tools related to payments or transfers are registered in the code, ensuring that sensitive operations cannot be completed without user confirmation. All dialogue history and permission records are stored locally, with only current task information transmitted to the cloud; inference results are discarded immediately after completion, with no persistent storage. These safety designs aim to maintain human control over AI in complex execution scenarios.

Item Li Gang, Chairman of the Zhongguancun Information Consumption Alliance, told Beijing Business Daily that the core advantage of hardware manufacturers deploying native AI agents lies in tightly binding capabilities with hardware—controlling system permissions, computing power, and ecosystems to create a more controllable environment, avoiding vulnerabilities of proxy modes, and securely managing and invoking smart devices.

The development of native AI agents in hardware requires aligning with technological and market drivers, and finding a practical balance between capability expansion and risk prevention that fits industry realities.