Build a Practical AI Agent from Scratch with Coze’s Low‑Code Platform

Some say that discussing large models without mentioning agents means thinking is stuck in 2024, and there is some truth to that. The 2025 World AI Conference highlights agents such as digital‑human agents, procurement agents, and live‑broadcast agents. This article shows how an individual can build a practical agent from zero.

Coze Platform

Among several agent‑building platforms, the author chooses Coze , a low‑code/zero‑code development platform launched by ByteDance. It supports multimodal inputs (text, voice, image) and a visual workflow designer, allowing users without deep programming skills to create agents. Finished agents can be published to Douyin, Feishu, Doushan and WeChat public accounts.

Agent Construction Logic

The overall process includes:

Register and log in to the platform (国内 https://www.coze.cn/ 国外 https://www.coze.com/).

Define the agent’s goal (customer‑service assistant, e‑commerce recommender, personal knowledge‑management helper, etc.).

Configure the workflow – the core step, done by dragging and linking nodes.

Create the agent and publish it.

Workflow Construction

After registration, select the workspace → resource library → resources → workflow:

工作空间-资源库-资源-工作流

The example builds a "Public‑Account Writing Assistant" agent. First, give the workflow a name (letters and numbers) and describe its purpose.

The workflow starts with a 开始 node and ends with an 结束 node. Intermediate nodes are added by dragging from the 开始 node.

Edit 开始 Node

Set the variable name (e.g., a01) and type (String for text input).

Design Text‑Writing Node

Add a large‑model node (e.g., Doubao • 1.5 • Pro • 32k ) and configure skills and knowledge base.

Input variable a01 from the 开始 node is renamed a02 for the text‑generation step, and the output variable is b02 (String).

Prompt design uses two types of prompts:

System prompt : provides system‑level guidance such as persona and response logic.

User prompt : gives the model the user’s instruction.

根据主题{{a02}}写一篇2000字左右的文章

Design Image‑Generation Node

Add an image‑processing → image‑generation node, choose a model, set image ratio, quality, reference image, and prompt.

Edit 结束 Node

Configure the final outputs: text from b02 and image from data.

Test Run and Publish

Run the workflow with a sample theme (e.g., "数学"). The result shows both article text and cover image.

After confirming the output, click the 发布 button and authorize the desired platforms.

Create the Agent

Return to the workspace, choose 项目开发-创建智能体, set the agent’s name, description, and icon.

Add the previously designed workflow, optionally set a background image and opening speech, then save.

Now the agent can converse and generate both text and images.

After publishing and authorizing the platforms, the agent is ready for use.

Agents have broad applications—from personal assistants to enterprise customer service and e‑commerce recommendation systems—greatly improving efficiency and user experience. The low‑code nature of Coze makes it easy to customize agents for various domains.

Don’t hesitate, try building your own agent now!

If you found this article helpful, please like and share.

For those interested in the fundamentals of large language models, the author recommends two books (titles omitted for brevity).