HarmonyOS On-Device Assistant Upgrade: Multi-Agent Collaboration and Voice-Driven PPT Generation Explained

Mobile AI assistants are evolving from single-turn Q&A tools into task-oriented agent platforms capable of scheduling, collaboration, and execution. A recent major update to a HarmonyOS on-device assistant focuses on three areas: PPT generation with custom layout templates and voice interaction, an Agent cluster for parallel multi-agent scheduling, and a skill invocation mechanism that connects with common office and communication applications. These capabilities are not superficial UI refinements. They reflect engineering approaches to task decomposition, concurrent execution, and cross-app coordination in on-device AI.

This article examines the design logic, applicable scenarios, and practical usage of these features from a technical perspective, helping developers and heavy office users understand the underlying mechanisms.

1. PPT Generation Upgrade: From Generating Results to Understanding Intent

PPT generation tools built on large language models have long faced a common problem: the model can produce a well-structured deck, but it struggles to grasp the user’s actual intent. The vaguer the description, the more likely the output diverges from expectations. This update addresses the pain point with two improvements.

1.1 Custom Layouts and Templates: Constraining the Generation Space in Advance

The core value of custom layout templates lies in letting users define output boundaries in a structured way before generation begins. For example, a business report typically requires a structure of “conclusion first + data evidence + risk warning,” while an academic defense emphasizes “research background + methodology + experimental comparison + conclusion.” With a preset template, the model no longer searches blindly in a vast solution space; instead, it generates content within an established framework.

From an engineering perspective, this introduces stronger prior constraints during inference. A template not only defines page layout but also implies content organization logic and information density requirements. For developers, the “template as constraint” approach is worth adopting: when designing generative applications, offering a limited set of high-quality structured options often improves user satisfaction more than unrestricted free-form input.

1.2 Voice Interaction: From Command Input to Conversational Clarification

The addition of voice interaction has practical motivations. Mobile users find long text input inconvenient, yet PPT requirements often involve multiple layers of logical relationships. Through voice-call-style interaction, users can describe their needs naturally while the system completes information gathering and confirmation during the conversation.

The key to this capability lies in semantic understanding and task clarification after speech recognition. A well-designed conversational generation flow should have the following characteristics:

• Recognize priority and emphasis in the user’s tone;
• Proactively ask follow-up questions when critical information is missing, without interrupting the user’s expression;
• Convert multi-turn dialogue content into structured generation parameters.

For instance, if a user says, “Create a project retrospective for me, emphasize data growth, and keep the style clean,” the system needs to extract three key constraints — “project retrospective,” “data growth,” and “clean style” — and prioritize them during generation. If the user adds, “Add an executive summary at the beginning,” the system should map that instruction to adjustments in the first page’s content and layout.

For users of this feature, it is advisable to clearly state the audience, occasion, and style preference when describing requirements verbally. Clear requirement expression, combined with the clarification mechanism of voice interaction, significantly improves the usability of the first draft and reduces manual revision effort.

2. Agent Cluster: Decomposing and Parallelizing Complex Tasks

The core highlight of this update is the Agent cluster capability. Compared with a single agent completing tasks serially, the value of an Agent cluster lies in decomposing complex tasks into multiple sub-tasks that can be executed in parallel by multiple agent instances.

2.1 Limitations of a Single Agent

A single agent faces obvious bottlenecks when handling long-chain tasks. For example, writing an industry research report involves information retrieval, data organization, competitive analysis, insight extraction, and structural arrangement. If only one agent performs these steps serially, the process is not only time-consuming but also prone to context loss and logical gaps.

The Agent cluster approach assigns these steps to sub-agents with different functions. Each sub-agent can focus on a specific domain or step, such as:

• Retrieval agents responsible for collecting materials from multiple information sources;
• Analysis agents responsible for statistical processing and interpretation of data;
• Writing agents responsible for transforming analysis results into coherent text;
• Proofreading agents responsible for checking logical consistency and format compliance.

Multiple agents work in parallel during the same time window and then merge their results. Under this model, complex tasks that previously required extensive manual effort can deliver initial drafts in a much shorter time.

2.2 Task Scheduling and Result Merging

The engineering challenges of an Agent cluster lie in task scheduling and result merging. Task scheduling must address questions such as “how many sub-tasks should the overall task be split into,” “what are the dependencies between sub-tasks,” and “which sub-tasks can run in parallel versus sequentially.” Result merging must ensure that output from different agents remains consistent in style, terminology, and logic.

From a practical standpoint, a reasonable Agent cluster workflow typically includes the following steps:

User submits a complex task
       |
Main control agent analyzes and decomposes the task
       |
Generates a sub-task list and dependency graph
       |
Assigns sub-tasks to specialized agents based on type
       |
Parallel execution (dependent sub-tasks trigger sequentially)
       |
Collects output from all agents
       |
Main control agent consolidates, deduplicates, and unifies style
       |
Returns the final result

For ordinary users, the above process is encapsulated behind a simple action: switching to Agent cluster mode. The system automatically determines the sub-task decomposition strategy based on task type, with no manual intervention required.

2.3 Applicable Scenarios

The Agent cluster is especially suitable for the following types of tasks:

• Literature reviews and academic theses: reading large volumes of literature, extracting key arguments, categorizing by theme, and generating synthesis text;
• In-depth industry research: collecting information from multiple dimensions, including market size, competitive landscape, technology trends, and policy environment;
• Multi-document comparative analysis: processing multiple documents in parallel and extracting differences and common characteristics;
• Long report writing: generating initial drafts for each chapter before unifying logic and writing style.

In these scenarios, the user’s value lies in defining task objectives and reviewing final results, while the intermediate work of information collection and initial drafting can be efficiently handled by the Agent cluster.

3. Skill Invocation and Cross-App Collaboration

This update also introduces a skill invocation mechanism. The mechanism allows the assistant to call a large set of pre-built skills and establish connections with common communication and office applications. This means users can orchestrate cross-app operations within a single conversation interface.

3.1 Design Logic of Skill Invocation

The essence of skill invocation is encapsulating common operations as reusable functional modules. Examples include:

• Information retrieval skills: quickly searching for specific types of data or news;
• Document processing skills: format conversion, content summarization, keyword extraction;
• Communication and collaboration skills: message sending, notification reminders, file sharing.

These skills are exposed to the assistant through a unified invocation interface, so users do not need to switch back and forth between applications. Architecturally, this design treats the assistant as an orchestration layer, with underlying skills serving as execution units. The user expresses intent, and the orchestration layer selects the appropriate skill and completes execution.

3.2 Connecting with Common Applications

An important capability of the skill invocation mechanism is binding with common chat and office applications. Once bound, users can trigger cross-app operations through voice or text commands. For example, a user can ask in a conversation to send a document summary to a specific contact, and the system automatically completes both summary generation and message delivery.

This cross-app collaboration capability is particularly important for mobile users. Touchscreen devices involve longer operation chains than desktop environments, so any capability that reduces app switching delivers clear efficiency gains. For enterprise users, cross-app automation also lowers repetitive communication overhead.

3.3 Extensibility Considerations from a Developer Perspective

From a developer’s perspective, the opening of the skill invocation mechanism means more third-party skills can be integrated. If a skill exposes a standardized interface, any service conforming to the interface specification can theoretically be invoked by the assistant. For internal enterprise development, this offers a path to combine self-built services with an on-device AI assistant:

Internal enterprise service → wrapped as a standardized skill interface → registered to the skill invocation layer → invoked by the assistant on demand

This architecture positions the AI assistant not merely as a conversational tool but also as a human-machine interaction entry point for internal enterprise services. It should be noted that permission control and data security deserve close attention during integration to prevent sensitive operations from being triggered without authorization.

4. Usage Recommendations and Efficiency Optimization

For users who want to fully leverage the above capabilities to improve productivity, the following practical suggestions may be helpful:

4.1 Define Task Boundaries and Use the Right Mode

For everyday simple Q&A and quick information retrieval, regular conversation mode is sufficient. When facing complex tasks involving multi-dimensional information collection, long-form text generation, or deep analysis, switching to Agent cluster mode is recommended. Matching the right tool to the right task is the first step in efficiency optimization.

4.2 Prioritize Templates and Structured Descriptions in PPT Generation

Precise template selection can significantly reduce subsequent revisions. Before generation, it is advisable to clarify the following:

• Presentation occasion: internal reporting, external proposal, academic presentation, or training session;
• Audience profile: management, technical peers, clients, or general staff;
• Content emphasis: conclusion-first or process-oriented presentation;
• Style preference: clean, professional, creative, or data-intensive.

4.3 Leverage Voice Interaction to Reduce Input Burden

In mobile scenarios, voice input is often more efficient than typing. This is especially true during the early stage when requirements are still unclear and users need to think aloud. Voice interaction captures ideas more naturally, and the system helps clarify requirements progressively through dialogue.

4.4 Convert Repetitive Work into Skill Invocations

For fixed operations performed regularly — such as sending weekly project status summaries or periodically retrieving industry news — it is worth identifying the rule-based steps and automating them through the skill invocation mechanism. This frees up more time for work that requires judgment and creativity.

5. Summary and Outlook

This HarmonyOS on-device assistant update delivers substantive enhancements across three dimensions: PPT generation experience, multi-agent collaboration, and cross-app skill invocation. The common direction of these capabilities is moving the on-device AI assistant from answering questions to completing tasks. Custom templates and voice interaction improve the efficiency of human-machine intent alignment; the Agent cluster addresses decomposition and parallel execution of complex tasks; and the skill invocation mechanism provides an engineering path for cross-app collaboration.

As on-device computing power grows and multi-agent scheduling technology matures, mobile AI assistants are expected to take on substantive work in more professional scenarios. For developers and enterprises, understanding the interaction patterns and integration approaches of these capabilities in advance can support more informed architectural decisions in AI application deployment.