I Made a Text RPG to Figure Out What Swarm Actually Is

OpenAI dropped Swarm this month — quietly, with a disclaimer that it's experimental and not for production, which is the academic paper equivalent of saying "don't try this at home" on a YouTube video where someone builds a flamethrower.

The core idea is almost insultingly simple. You have agents. Each agent has instructions and tools. Agents can hand off to other agents. That's it. No graph topology to configure, no YAML the size of a dissertation, no abstraction layer wrapping an abstraction layer wrapping a function call. You can read the entire source in an afternoon and feel vaguely suspicious that something is missing.

So I built a text-based RPG.

Not because it's a practical use case — it isn't — but because an RPG is a perfect stress test for exactly what Swarm is doing. You have a narrator agent who runs the world, a combat agent who takes over when swords come out, a merchant agent with its own personality and inventory logic, a dungeon master who hands you back to the narrator when the fight ends. Every handoff is a scene change. Every context variable is the party's health or gold or the fact that you already killed the innkeeper so you can't go back there.

The handoffs are the thing. Watching control pass cleanly from one agent to another — each one dropping into its role with its own instructions, its own tools, no global state getting tangled up — it does something to your understanding of what these systems want to be.

Whether Swarm is a prototype of something real or just a teaching tool they shipped because the internal demos were fun, I genuinely don't know. The "not for production" warning is the most important line in the readme. But the mental model it installs is useful in a way that no amount of reading about agents would have been.

Build something stupid with it. That's my advice.