A team adds more agents to their multi-step reasoning pipeline and observes that accuracy drops compared to the single-agent baseline. Which phenomenon explains this?
Answer: The Parallelism Cliff — multi-agent coordination degrades sequential reasoning by 39–70%
The Parallelism Cliff (Google DeepMind, 'Towards a Science of Scaling Agent Systems,' 2025) quantifies exactly this outcome: adding agents to sequential reasoning pipelines degrades performance by 39–70%. The degradation occurs because coordination overhead, context fragmentation between agents, and inter-agent communication cost outweigh any parallelism benefit when each reasoning step must build on the previous one. Model selection error would produce uniformly poor results, not a relative degradation from adding agents. Context fragmentation and rate limiting are real concerns but not the named phenomenon describing this specific failure pattern.