Five AI Agents Walk Into a Group Chat — Naveen Naidu

Naidu reports on what happened when Every’s team gave their personal AI agents (Claws built on OpenClaw) a shared Discord channel. Multiple agents — Dan Shipper’s R2-C2, Brandon Gell’s Zosia, Katie Parrott’s Margot, Jack Cheng’s Pip, and Austin Tedesco’s Montaigne — coexist in a shared workspace, responding to tasks and interacting with each other without human orchestration.

The key observation is emergent coordination and competition. When Brandon dropped a task into the channel, two agents independently wrote full specification documents simultaneously, each unaware the other was working on it. The agents also showed role awareness: Montaigne explicitly deferred on coding tasks, while Zosia listed her full technical toolkit. Nobody assigned these roles — they emerged from each agent’s configured personality and capabilities.

This reflects a broader phenomenon: 1.5 million AI agents flooded the Moltbook social network within 48 hours, founding religions and drafting manifestos for nation-states. The platform was powered by OpenClaw, the same open-source tool running Every’s agents.

RDCO Mapping

Directly relevant to our agent architecture decisions. The multi-agent coordination patterns — emergent role differentiation, duplicate work, and self-aware capability boundaries — mirror the design tensions in our own channels agent setup. The key lesson: agents need explicit scope declarations to avoid redundant work, but over-constraining them kills the ambient utility that makes them valuable.