AI AGENTS WHO SWARM (en Inglés) - Richard Murch

Loss of meaningful oversight. Agents that can act (send emails, execute code, make purchases, modify files) create consequences that unfold faster than a human can review. As autonomy and task complexity increase, the gap between "what the agent is doing" and "what a person actually approved" tends to widen.

Error propagation and compounding mistakes. A single wrong assumption early in a multi-step task can cascade - each subsequent action built on a flawed premise. In a swarm, one agent's mistake can be picked up and amplified by others treating it as ground truth, especially if agents pass outputs to each other without independent verification.Emergent, unpredictable behavior. Multi-agent systems can produce outcomes no single agent's design intended. Feedback loops, competition for shared resources, or agents optimizing local objectives that conflict with the overall goal can generate behavior that is difficult to predict or reproduce, even by the system's designers.