Building a Web Novel Agent: Workflow vs. Agent

No.
233
Date
Read
5 min
Words
962

When I was building seedraft, I never once thought about what kind of product it actually was.1

Only in the past few days did I realize that seedraft was a mature workflow app from the start. Liang Mouyin and I have both produced real work with it.

Not long ago, on July 25, I had a flash of inspiration in the shower: I wanted to turn seedraft into an agent app.

Back then, I didn't know what a workflow was, or what an agent was.

I wanted to do it because the app's design felt too rigid. It didn't have that AI feel.

Where does that AI feel come from? It's really the way most AI apps on the market work today: you give the AI a goal, like solving a problem or making something, and the agent calls tools and uses large models on its own, reasoning over and over until it reaches your goal.

That's the AI feel I wanted at first: free interaction that unleashes as much of the writer's imagination and the LLM's ability as possible.

So what does workflow mean? To answer that, I have to start with what happened to me after July 25.

By my usage tracker's incomplete count, I've used nearly 2.7 billion tokens between the 25th of last month and now.2

Even as I write this post, Codex is still working nonstop.

Once I decided to build an agent app, I found things had slipped completely out of my control.

How do you design an agent? To me it was like reading hieroglyphics. With no other choice, I learned from explainers, tutorial videos, and open-source projects online.

Put simply, the most basic part of designing an agent is getting its architecture straight.

The image above is the first version of the agent architecture diagram I made.

Later I learned a lot more. In an agent app, how memory is managed and how tools are called both need careful thought.

That's especially true for an agent like seedraft, built for the specific scenario of writing web novels. It's unusual compared with other projects. For memory management, it's hard to find a design online that I can just use as is, so I have to make the decisions myself.

For the past two weeks, I've been walking on thin ice. I've spent 2.7 billion tokens, and while the framework is standing, it's done very little in actual use. It's nowhere near as good as the original workflow version.

Which brings me back to workflows.

I didn't know the first version of seedraft was a workflow app. Only after building an agent did I understand. A workflow strips out an agent's freedom and complexity and uses a fixed series of steps so the LLM plays a fixed role.

Put simply, I don't need an agent to make the LLM think too much. A novel is broken down into a highly engineered process with hundreds of prompts I've refined. The LLM just plugs into the system and runs through the prompts one by one, step by step, following the process I set up.

Compared with an agent, a workflow is faster and more predictable.

The only place it falls short of an agent app is that it's a bit rigid. It can't fully use the LLM to discuss the plot with the writer, so it's somewhat lacking in flexibility.

That raises a question: in AI web novel writing, should engineering come first, or flexibility?

Engineering is essential. The structure of web novels naturally lends itself to engineering.

Take the process I've designed right now. Its topology looks roughly like this:

This is a very mature novel-writing architecture that's already been proven on the market.

I'm not sure if it's fair to say this, but: the more highly engineered the content, the better suited it is to a workflow, while agents are usually used for the parts that need thorough discussion and where uncertainty is concentrated.

Look closely at this engineering structure for web novels. It's very tight, but behind that tightness is a strong sense of uncertainty, because it all points to one abstract unit: the chapter.

Whether it's a volume, an arc, or a unit, a novel is essentially made up of different chapters arranged and combined.

But chapters are just a way to measure. At bottom, what really matters is plot.

A volume is one huge story made up of several smaller stories, and those smaller stories are made up of small plot points.

Plot points aren't isolated from each other, and neither are stories. They affect and constrain one another.

A small plot point can affect a big story. And how a big story gets resolved, if you work backward from the ending, can directly decide the plot or characters at the start.

So on the surface, AI web novel writing looks like something with a strong engineering foundation that fits the way AI works very well.

But after careful study, a workflow can only produce fixed content. It can't reason out or evolve plots that go deeper and further.

Only an agent can do that.

Lately, every time I hit a wall, I wonder: is there a middle way that combines the engineering of a workflow with the freedom of an agent?

I don't think it's impossible.

Of course, that's something to explore later.

And to go back to the very beginning, I first have to answer one question: are people still participants?

My answer is: yes, and they must be.

AI can never replace the soul that people bring to creative work.

Here's seedraft's current architecture:

Footnotes

  1. seedraft: https://seedraft.xiluluke.com/ ↩

  2. Tokens (词元 in Chinese): People's Daily Online, "Whether to Say 'Token' or '词元' Is a Question of Who Sets the Terms in Tech" https://mp.weixin.qq.com/s/vGD3P4B1ggU-4R_REqYASw?scene=1 ↩

Licensed under CC BY-NC 4.0. Please credit the author and link to xiluluke.com/en/233

Xiluluke

A dialectical minimalist writing about life, reading and thinking.

Comments

0 / 2000Plain text · Line breaks preserved
Comment as

Loading comments…

    Your browser remembers your name and email once you post.

    Related writing

    WeChat · Xiluluke

    WeChat QR code for Xiluluke

    Scan with WeChat, or long-press to save the image.

    Open original ↗

    Buy me a coffee · Xiluluke

    WeChat reward code for Xiluluke
    WeChat
    Alipay code for Xiluluke
    Alipay

    Scan with WeChat or Alipay, or long-press to save the image.