Princeton NLP: Single Agent Matched or Outperformed Multi-Agent Systems on 64% of Benchmarked Tasks — 40% of Multi-Agent Pilots Fail Within Six Months

Augment Code·medium signal

Princeton NLP research found that a single agent with the same tools and context matched or beat multi-agent setups on 64% of benchmarked tasks, and 40% of multi-agent production pilots fail within six months. The practical implication for builders: before investing in multi-agent orchestration complexity, verify your task actually benefits from it. The research suggests the orchestrator-worker pattern (capable model orchestrates, cheap models execute) cuts costs 40-60% when parallelism genuinely helps, but most tasks don't need it.