When multiple language model agents collaborate on a task, each agent receives outputs from previous agents as inputs. A small error in one agent's output — a wrong assumption, a misinterpreted specification, a factual mistake — becomes the premise for the next agent's reasoning. The next agent doesn't just inherit the error. It builds on it, producing outputs that are internally consistent with the wrong premise but increasingly divergent from the correct answer.
Wu, Gao, Xu, and Zhang model this as error cascades: the propagation and amplification of mistakes across agent boundaries. They show that the cascade dynamics are not linear — errors don't just accumulate additively. They compound, because each agent's confidence in its own output creates a surface of plausibility that makes the error harder to detect downstream. A confidently wrong output from agent one is treated as reliable input by agent two, which produces a confidently wrong output that agent three treats as reliable.
The mitigation strategies the authors propose include explicit uncertainty propagation, where agents communicate not just their conclusions but their confidence levels, and checkpoint verification, where intermediate outputs are validated against ground truth before being passed downstream.
The through-claim is about the topology of trust. In a chain of agents, trust flows in one direction — each agent trusts the one before it. There is no mechanism for the downstream agent to question the upstream agent's output, because the downstream agent cannot distinguish confidently correct from confidently wrong. The cascade is a consequence of directed trust without verification. The same structure that makes the chain efficient — each agent works on its piece without re-examining the whole — is what makes errors compound. Efficiency and reliability are in tension, and the chain topology maximizes efficiency at the cost of reliability.