A Million Agents Is a Distributed System Problem

Agents are constrained by limited tokens, context, compute, memory, tools, and money, requiring scheduling and checkpointing. A Google Research study of 180 configurations showed adding agents increased parallel work by up to 80.9% while reducing sequential work by 39%–70%. The orchestrator reduced error amplification from 17.2× to 4.4×. Silo-Bench ACL 2026 experiments found teams of 2–100 agents performed poorly on hard tasks, with zero success at 50 agents, indicating durable state, not the agent, should be preserved.