Error amplification describes a situation in which mistakes become more frequent or consequential as a task is passed through several AI agents. A wrong assumption from the first step may be repeated, expanded or acted upon by agents that never inspect the original evidence.
It is useful to compare a multi-agent process with a single-agent baseline under matched tasks and evaluation rules. Researchers may examine final accuracy, the number of intermediate mistakes or the impact of injected faults. The exact amplification measure must be defined because no single universal formula covers every study.
Not all multi-agent systems amplify errors. Independent verification, diverse reasoning methods or well-placed human review can catch mistakes and improve outcomes. Coordination overhead and pressure to accept another agent's output can have the opposite effect.
To measure amplification, record step-level outputs and compare final outcomes with and without agent handoffs. Test both ordinary tasks and deliberately corrupted intermediate messages, while holding tools, budgets and scoring as constant as practical.