Exploring AI Agent Integration in HandCat App: Architecture, Tool Management, and Implementation

With the development of large language models (LLMs), AI agents have become a hot topic. This article summarizes the HandCat team's exploration of AI agents, technical challenges, and practical implementations.

Agent is defined as LLM + memory + planning + tool use. The decision process includes perception → planning → action.

Key components of an LLM‑driven agent: planning, memory, and tool usage. Tools can be external APIs, databases, etc.

The HandCat project builds an end‑to‑LLM communication pipeline, session management, and tool abstraction. The architecture separates the agent template, tool protocol, and view layer. The agent returns two structures: object for LLM consumption and viewDTO for the app display.

Tool abstraction defines three types:

General tools – non‑interrupting, executed directly.

Selector tools – interrupt the flow to let the user choose an item.

Interruptor tools – pause execution for external input.

Examples of JSON responses:

{
    "code":200,
    "success":true,
    "object":[{"itemId":665788007683}],
    "viewDTO":{"sceneName":"showOrder"}
}

When an error occurs, the response may be:

{
    "code":-1,
    "success":false,
    "message":"没有权限",
    "viewDTO":{"sceneName":"showError","showTips":"抱歉，你暂时没有权限"}
}

Tool granularity affects performance: larger granularity reduces round‑trips but lowers reuse; smaller granularity increases flexibility but adds latency.

Safety considerations include parameter passing, validation, idempotency, and transaction handling.

Challenges observed: high correctness requirement for C‑end users, LLM output stability for structured UI data, tool‑LLM interaction stability, and token consumption.

The HandCat team built an “Agent Lab” to let developers quickly prototype agents, combine tools, and test end‑to‑end flows.

In summary, the project demonstrates a practical pipeline for integrating LLM‑driven agents into a commercial mobile app, covering architecture, tool management, error handling, and safety.