AI Intelligent Agent Preliminary Exploration

Title: A First Look at AI Agents

Introduction

For a long time, computers hardly possessed any intelligence; programming languages were the only way to control them. Ordinary people had to rely on applications built by engineers to translate user intent into machine capabilities. With the rapid development of AI technologies led by large models such as GPT and DeepSeek, computers can now understand human language and execute commands according to predefined rules and workflows. Natural language is becoming the standard way to control computers, and programming applications are no longer the sole option. In many scenarios, simply referencing templates and examples to write prompts, skills, or assemble an agent can solve the problem.

https://en.wikipedia.org/wiki/DIKW_pyramid

Looking at the DIKW model chain Data → Information → Knowledge → Wisdom https://baike.baidu.com/item/DIKW%E4%BD%93%E7%B3%BB/9860192:

Traditional computers: operate only up to the information level, lacking the ability to understand and apply knowledge.
Intelligent computers: have reached the knowledge level and can begin to use knowledge autonomously to solve problems.

AI is changing the human‑computer interaction paradigm: humans adapt to machines → machines adapt to humans

NNgroup: First New UI Paradigm in 60 Years

Each technological breakthrough enables more people to use computers to solve more problems, and this AI wave will be even more far‑reaching. Therefore, we need to consider how to make the best use of “intelligent machines,” leveraging human knowledge and wisdom to tackle harder challenges and improve life.

Core Capabilities of AI

Why can AI work like a human? In my view, it possesses several abilities:

Extraction: pulling valuable information according to a specified intent, much like a person researching sources.
Compression: merging and distilling key points, similar to human summarization.
Expansion: supplementing and fleshing out content on a topic, akin to a person elaborating a viewpoint.
Imitation: rewriting according to a given template, just as a person copies a model essay.
Generalization: applying existing knowledge to new problems, like a person drawing analogies.

These are also typical workflows for knowledge workers and form the basis for AI’s ability to write articles or code.

In addition to these core abilities, AI engineering platforms expose the following functions:

Cloud Services: quickly deploy a cloud‑based service.
OpenAPI: call third‑party APIs to fetch data.
MCP: allow AI to use third‑party services via standard protocols.

With these, users can build more powerful agents—or even full applications.

What AI Excels At

This is a hot topic right now. Recent insightful articles include:

From my understanding and practice, the following passage from the second article is fairly accurate:

Because what we call “high‑level work”—coding, analysis, charting, report writing, drafting contracts—is essentially the manipulation of symbolic information. AI’s greatest strength is precisely handling symbolic information. A legal assistant may need three days to retrieve case law; AI does it in three seconds. A junior programmer might spend a day writing code; AI finishes in a minute. A financial analyst could spend half a day on Excel; AI completes it in a fraction of a second. These tasks sound “high‑end,” but they share clear rules, well‑defined patterns, and abundant data—exactly the domains where AI shines.

Any work that fits these characteristics—applying existing knowledge from a specific perspective within a defined framework and process—can be assisted by AI. Complex computation and reasoning, however, remain beyond its current capabilities.

Getting Started Quickly

If you are a casual user who just wants to ask questions, tools like Qianwen, Doubao, DeepSeek, etc., are already sufficient. If you want AI assistance for writing and knowledge management, consider YouMind. If you aim to provide services with AI, you’ll need to build an agent. Currently, three platforms stand out:

Alibaba Cloud Bailei: https://bailian.console.aliyun.com/ https://bailian.console.aliyun.com/
Tencent Yuanqi: https://yuanqi.tencent.com/ https://yuanqi.tencent.com/
ByteDance Coze: https://code.coze.cn/ https://code.coze.cn/

Bailei targets enterprise users and offers a relatively pure agent‑development environment, though it lacks convenient WeChat integration and is best used via API. Yuanqi’s strength lies in its WeChat ecosystem; building agents from public‑account content is a key feature, and the resulting agent can be accessed through a mini‑program, though the building experience is average. Coze is aimed at the mass market, trying to create an AI‑era workstation. It updates rapidly and packs many features; the agent‑building part is just one capability, and the overall experience can feel a bit cluttered, though the platform itself is powerful.

Based on our testing, we recommend the following platform choices:

1. For simple agents within the WeChat ecosystem, choose Yuanqi first. Remember to select the DeepSeek model and set the temperature above 1.0; otherwise the responses can be too rigid.

Example: “Energy Boost”

2. For agents with higher demands, start with Coze. The entry point is [here]; just follow the product’s workflow to complete the build.

If you want to deploy a Coze‑built agent as a mini‑program, you can use the “Xiaowei Agent” option and follow the guide—setup is quick.

Related documentation:

Example: “Miaoli” (scan the QR code in WeChat to use)

Writing Prompts & Training Agents

Sample Template

Imitation is the best form of learning.
First, define the agent’s persona and clearly describe what you want it to do. Then use “Auto‑Optimize Prompt” to generate a template, which you can tweak as needed.

The prompt uses Markdown syntax, a markup language widely used by engineers for documentation. This syntax also helps control AI output, so a brief learning session is worthwhile—see the resources below:

Besides “Auto‑Optimize Prompt,” Coze’s “Prompt Library” offers tutorials on prompt writing.

For deeper study on crafting high‑quality prompts, see:

Typical Structure

From the output of “Auto‑Optimize Prompt,” a prompt generally includes:

Role
Goal
Skills
Workflow
Output format
Constraints

In practice, an agent is usually backed by a guiding philosophy or theory that reflects the creator’s wisdom. Adding a section to describe this can be helpful:

Thinking Framework: Empowering AI with wisdom

A Case Study

(Insert case details here)

Training Insights

Core principle: clear rules, obvious patterns, abundant data

1. Define the knowledge scope: AI has vast knowledge, but we should constrain it to reliable sources—textbooks, classic works of each discipline—to boost credibility and reduce hallucinations.

2. Choose the right model: For professional‑level services that demand strong comprehension, pick DeepSeek. For more casual, lifestyle‑oriented tasks, Doubao works well.

3. Outline the thinking framework: Humans naturally follow a mental roadmap when answering questions. To get the AI to respond as expected, formalize that roadmap into a clear framework for the AI to follow.

4. Use structured instructions: Directing AI isn’t like giving a person vague directions; it requires a structured, step‑by‑step set of rules.

Managing Existing Knowledge with a Knowledge Base

If you want your agent to use proprietary data, store it in a knowledge base and present it in a format that AI can easily process.

Start Building Your Product

Everyone harbors a product dream to some degree; AI brings that dream closer. Building an agent can be the first step toward realizing it. From a spark of enthusiasm to a full‑blown business, the journey often looks like: Passion → Idea → Prototype → Product → Business. Within this flow, product thinking is crucial—it’s the core mindset of internet companies. I recently saw an unexpected discussion of product thinking in an article by a traditional Chinese medicine practitioner:

It’s an entertaining read that gives a glimpse of this mindset. For deeper learning, check out WeChat’s own resources:

Zhang Xiaolong’s “Product Philosophy Behind WeChat” – Transcript

2026.03

Originally written by Ping Xia (平侠) and published in Chinese on 研习录 (Study Notes). Translated and adapted for DriftSeas with permission.