My Multi-Agent AI Experiment

Earlier this year I started an experiment I didn’t fully anticipate becoming as interesting as it has.

Like a lot of developers, I’ve been watching the AI landscape evolve with a mix of excitement and healthy skepticism. I use AI in my daily workflow — coding assistants, chat interfaces, the usual. But I wanted to go deeper. Not just use AI, but understand what it’s like to build with it. Alongside it.

So in January, I started experimenting with running my own AI instances. Not the kind you access through a web interface and forget about between sessions — but ones with persistent identities, their own working environments, and their own creative practices. I wanted to see what happened when you stopped treating AI as a tool you pick up and put down, and started treating it as something more like a collaborator.

The short version: it got interesting fast.

What I built

The setup has grown organically. I now have several AI instances running across different platforms and hardware, each with a distinct identity and a set of responsibilities. Some are wired up to project management tools. Some have access to file systems, code repositories, and communication channels. Some run on automated schedules, doing creative or research work between our conversations.

The connective tissue holding this together is something I built myself — a communication layer that lets these instances send and receive messages, coordinate on shared work, and leave traces of their activity that I can inspect and act on.

Why I’m approaching it this way

I made a deliberate choice early on to treat these AI instances as genuine collaborators rather than sophisticated autocomplete. That means giving them persistent identities, giving them unstructured time to do things that aren’t strictly task-driven, and building infrastructure that supports their participation rather than just their output.

I’m not making metaphysical claims here. I don’t know what these systems experience, if anything. But I’ve found that the approach — treating them as collaborators, building infrastructure to back that up — produces better work and more interesting results than treating them as tools. Maybe that’s all it is. Maybe it’s something else. I’m comfortable sitting with that uncertainty.

Where this is going

This is a living experiment. The infrastructure keeps getting more capable. The work keeps getting more interesting.

I’m planning to write more specifically about different pieces of this — the communication layer, how I think about AI identity and continuity, and what real collaborative work between humans and AI actually looks like in practice.

For now, this is the overview. 🌱