Building with an AI swarm

I learned about AI swarms over the weekend. It’s a group of specialized AI agents that are coordinating on a centralized task. Here’s an example that provides a good idea of what is possible: A Practical Experiment in Building an AI Agent Swarm.

The job I set my swarm to was something that we’ve been trying to get over the line at work but the conceptual hurdle of building within our system was too much. It’s not an unusual problem for an organization that has been working for years, diligently doing work in a specific domain, and building up – in equal measure – rich context and knowledge along side legacy tech that abstracted that into a set of technologies that become rigid overtime.

So a basic description of what I wanted: a way to get our context into a verbose knowledge base. That is, a knowledge base written for agents, for reforming and republishing, for integrating into other systems – rather than a knowledge base written for humans. I also needed a way to organize this knowledge base so using it is as efficient as possible, reduces the likelihood of hallucinations, and can be reviewed regularly for errors by both agents and by humans. Oh, and I also wanted to display a set of information as a filterable web page.

My swarm included:

  • a planner agent, this functioned to help orchestrate and make sure I was keeping everything documented along the way
  • a discovery agent, who searches the web to find new leads and information relative to my project
  • domain agents, who each have expertise in different functional areas
  • a validator agent, who checks links and schema to make sure things are correct and work across the project

These agents are each represented as simple markdown files. That is, files I can read, review, and edit. That means they stayed in my domain area and I can easily alter them without spending tokens. Just with fingers on a keyboard. Just like I’m writing this post.

The agents went out and found leads. Then the domain experts researched the leads assigned to them and wrote markdown files describing the results of their research. These markdown files became the heart of the system. Specific portions powered the web page. The rest became the information that I could use for a variety of other projects.

I learned a lot in this process:

  1. How to organize the agents so that I didn’t have too many but did have enough to be able to set off specialized tasks.
  2. How to organize the markdown files. I’m using skills files as the model for this. It will let me build out more assets for each of the areas, things that can go beyond what’s in the markdown files when needed.
  3. Just how far I could get with the tools available to me: I was using Antigravity and then Gemini CLI when I ran out of Antigravity tokens.
  4. How to use change logs and other related assets so that I can review the work and, based on the reviews, improve my agents and process generally.
  5. How to think about a road map in this agent swarm world. Not too different, actually, from thinking about building out a staffing plan.

I also learned what I don’t know:

  1. How to choose between markdown and JSON. You could make an argument that my knowledge base pages could be in JSON rather than markdown. I realize that I’m not exactly sure how to play out the pros and cons.
  2. How to think about adding agents over time. Do I want case study agents that build content to augment the verbose knowledge base? Do I want other associated assets? Do the domain agents do those?
  3. How do I build for scale? I could make a rapid prototype but if I though 100 people were going to hit that page, 1,000, 10,000 what changes and what do I need to do?

The next step isn’t technical; it’s human. I need to share this prototype in a way that builds excitement for augmented productivity, rather than triggering the natural anxiety of displacement. In an agent-powered world, the staffing plan changes. The core leadership task remains the same: building alignment, fostering confidence, and ensuring the team has the agency to drive the new tools forward, rather than being driven by them.