Single-Agent vs Multi-Agent Systems: Architecture Tradeoffs

Single-agent systems = monolithic LLM orchestration.
Multi-agent = distributed task execution with coordination layer.

Core difference

Single-agent:
input → prompt → output

Multi-agent:
input → router → agents → aggregator → output

Tradeoffs

Single-agent → low latency, low infra
Multi-agent → high flexibility, higher cost

Key insight

Most “multi-agent systems” fail because:

• no orchestration layer
• poor task decomposition

Practical rule

Start monolithic
Add agents when:
context window limits reached
tasks diverge