I Used an Agent to Build My Automated Workflow

What This Article Covers

Breaking "AI automation" into two concepts: an Agent workflow builds the process, while a programmatic automation workflow runs it on its own. Using a Codex-powered LINE backup bot as the example, I explain which parts to hand off to AI and which to hand off to code.

Separating the Two Concepts First

A lot of people lump automation together when they talk about it. I prefer to separate it into two concepts: an Agent workflow is responsible for building the process, while a programmatic automation workflow is responsible for running it once it is built. They require very different things.

A Real Example: The LINE Backup Bot

I recently used Codex to set up a LINE bot for myself. The goal was simple: every message in the group, timestamped, with sender and any attached files, automatically backed up to my computer, organized by group and project.

Once the LINE backup bot is configured, it is purely code running on its own:

This phase is programmatic automation. It has no direct relationship to AI and does not necessarily require a GPT API key. It just follows the configured rules and keeps running.

So Where Does AI Come In?

AI appears in the earlier phase, when the process is being built.

Rather than slowly researching LINE webhooks, permissions, how to wire up the code, and how to test it myself, I simply gave the task to Codex. That is how I now think about it: I used an Agent workflow to build a programmatic automation workflow.

Skills work on a similar principle. They lock down a way of doing something so it can be repeated the next time. That is the core idea behind a workflow: have you turned a task into a system that can be executed again and again?

Arranging these approaches from fixed to flexible, you can see a spectrum: hard-coded scripts, drag-and-drop workflow tools, Skills as reusable operating manuals, and fully autonomous Agents. The more fixed and repetitive the task, the more it belongs at the fixed end of the spectrum. The more variable and judgment-dependent it is, the more it belongs at the flexible end. The LINE backup bot sits firmly at the fixed end; Skills sit somewhere in the middle.

Push it one layer further: write the rules as plain-text documents, the way Skills are written, and you are no longer locked into any single AI. Use Codex today, switch to a different model tomorrow, and as long as it can read the document it can pick up where things left off. What truly persists is that reusable operating manual. Individual conversations come and go. The document stays.