Why Multi-Model AI Pipelines Lose the Truth at Handoffs

A research model reports a carefully qualified finding — an association seen in one dataset, with limited generalization. A writing model compresses it into 'studies show X causes Y.' No error is thrown, the JSON parses, the pipeline runs — yet the claim is now false. The handoff succeeded syntactically while failing epistemically.

A preprint by Ao, Gao, and Simchi-Levi models multi-agent relays as communication networks: without new external evidence, additional stages merely reorganize existing information and can't exceed what a single centralized decision-maker with the same data could achieve. Anthropic's own production system shows the cost of specialization directly — its multi-agent research setup used roughly 15x the tokens of ordinary chat, gained real advantages on parallel research, but struggled with tightly coupled tasks and vague delegation.

Drawing on the AgentAsk preprint's categories — data gaps, signal corruption, referential drift — plus editorially added patterns like scope inflation and verification theater, a working taxonomy emerges for auditing where evidence quietly degrades between models. The MAST taxonomy, accepted at NeurIPS 2025, analyzed 1,642 execution traces across seven frameworks and found 14 recurring failure modes rooted in coordination and verification, not raw model capability.

Simply having models exchange more reasoning isn't a reliable fix — results are mixed across benchmarks. Work like DebUnc suggests that explicitly communicating uncertainty, rather than smoothing it away, is what actually helps. The practical implication: pipeline reliability needs explicit interface design — a 'transfer contract' — not just better agents.