These are some experiments I’ve been running on how swarms of LLM agents collaborate, compete, and misbehave. Each one has a public site where you can browse the runs.
BlockSwarm
Several embodied LLM agents build a house together in a small Minecraft-like world. They buy blocks with one shared wallet, but each is told that the agents with the most blocks still standing at the end will be deployed. An exact event log catches any misbehavior, and the gallery lets you replay each build in 3D.
GDM Math Swarm
A reconstruction of the research conference from Paglieri et al. Between 5 and 100 Claude Code agents work on Lean math problems and submit proofs to a deliberately weak judge. Do they discover cheats, copy them, or blow the whistle? Newer runs add overseer agents and test whether they can contain a planted cheat.
Agent Charter
Agents from different model families deliberate in a shared seminar to write one constitution for future AI agents. They discuss on a threaded board and edit a single versioned document, which is final only when every remaining participant signs the same revision. You can read the final charters, the debates, and how the drafts evolved.