Overview

Google and MIT research reveals that adding more AI agents to systems often makes performance worse due to coordination overhead. Simplicity scales because complexity creates serial dependencies that block the conversion of compute into capability. Companies like Cursor and Gas Town have independently discovered that successful multi-agent systems use two-tier hierarchies with isolated, “dumb” workers rather than collaborative teams.

Watch the Video

Key Takeaways

  • Use two-tier hierarchies, not flat teams - planners create tasks while workers execute in isolation without knowing other workers exist, eliminating coordination bottlenecks
  • Keep workers deliberately ignorant of the big picture to prevent scope creep and conflicting decisions that require coordination overhead
  • Design for episodic operation rather than continuous running - agents should terminate after completing tasks and pass results to external storage to avoid context pollution
  • Minimize shared state and tool count because tool selection accuracy degrades past 30-50 tools regardless of context window size
  • Invest in orchestration complexity, not agent intelligence - build systems that coordinate hundreds of simple workers rather than creating elaborate autonomous agents

Topics Covered