I’m Building a Small Civilization of AI Agents

I've been experimenting with a slightly unusual idea: what if, instead of asking one AI to do everything, I gave different AI agents different jobs?

That's how my AI Civilization project started.

The basic idea is pretty simple. I want several small agents working together, with each one responsible for a particular part of a task.

For example, one agent can search for opportunities. Another can research them. Another can write something based on the research, and another can check the result.

It sounds a little like running a small company, except the "employees" are software.

Why not just use one AI?

I've used single AI prompts for a lot of things, and they work surprisingly well. But there's a problem when the task becomes bigger.

You end up with one huge prompt telling the AI to search, decide, write, check everything, and somehow remember what happened before.

That can get messy.

I'd rather split the work.

My current workflow looks roughly like this:

Scout → Researcher → Writer → Worker → Quality Checker

The Scout looks for opportunities.

The Researcher tries to figure out whether an opportunity is actually worth looking at.

The Writer prepares the content or application.

The Worker organizes everything into a package.

Then the Quality Checker makes sure we haven't produced something obviously incomplete or broken.

For now, there's still a human at the end of the process. That's intentional.

The part that's harder than I expected

Finding opportunities isn't actually the hardest part.

The internet is full of job boards and websites claiming to have remote jobs.

The difficult part is finding something that is:

• actually available,
• relevant to the person applying,
• open to their location,
• worth the effort,
• and something they can honestly do.

At first, my system was happily finding pages containing phrases like "10 Best Remote AI Jobs."

Technically, it had found what I asked for.

But it hadn't found a job.

That was a useful lesson.

I changed the system to use structured job listings as another source instead of relying entirely on general web search.

That immediately produced much better results.

AI shouldn't fill in the blanks

Another thing I've learned is that an AI writing an application can become dangerous if you let it guess.

If I don't have five years of professional copywriting experience, the system shouldn't suddenly decide that I do.

If I don't have a portfolio, it shouldn't invent one.

If a job requires experience I don't have, the system should tell me that.

That's why I'm experimenting with keeping a simple profile containing things I can actually verify.

The Writer can use that information when preparing an application, but missing information stays marked as missing.

It sounds obvious, but it's surprisingly important.

Where I want to take it

I'm still early in the project.

Right now I'm more interested in proving that the basic system can actually produce useful work than in building hundreds of agents.

Eventually I'd like the civilization to handle different kinds of work—research, content, data tasks, software-related work, and other digital services.

But first I want one part of it to work properly.

If I can get five real opportunities, prepare five honest applications, and learn from what happens afterwards, I'll have something much more useful than a fancy demo.

I'll have the beginning of a system that can actually learn from the real world.

And that's the experiment I'm interested in.