Multi-agent orchestration frameworks have quickly become the enterprise default for automating complex knowledge workflows. Modern platforms make it deceptively simple to chain specialized LLMs sequentially across distinct operational roles—research, quantitative analysis, drafting, and editorial review. Yet this linear architecture conceals an existential engineering flaw: downstream agents blindly trust their upstream peers without verifying primary sources, enabling early factual corruptions to compound silently into authoritative-sounding fiction.
In a research paper accepted at the FAGEN Workshop at ICML 2026, Prabhjot Singh of The University of Texas at Austin and Bhushan Pawar of RediMinds Inc. mathematically and empirically formalize this failure mode. They demonstrate that factual errors do not merely survive handoffs unchanged. Instead, mistakes actively mutate across sequential stages, transforming from explicit numerical discrepancies into high-level strategic narratives that evade standard auditing.
The Markov State Transition of Laundering Errors
Singh and Pawar frame this breakdown as the "hallucination snowball effect," modeling error propagation as a first-order Markov process spanning four states: Raw Fact, Derived, Narrative, and Invisible. In their experimental pipeline, an isolated numerical misstatement at Stage 1—such as recording cost of goods sold as $71.2B instead of $63.1B—is accepted as ground truth by an analytical agent at Stage 2 to compute a fabricated percentage shift. By Stage 3, a drafting agent embeds that metric into plausible qualitative prose, linking the distortion to macro supply chain disruptions. Finally, a review agent at Stage 4 evaluates solely for internal narrative coherence, rubber-stamping the fabricated claim.
"Raw numerical facts become derived computations, then narrative prose, then editorially approved conclusions. At each transformation, detectability degrades near-irreversibly."
This structural metamorphosis systematically destroys data provenance. Tracking 346 injected hallucinations across a four-agent financial workflow on FinanceBench, the authors measured an escalating escape probability at every boundary. The likelihood of an error slipping past unnoticed surges from 24.6% at the initial handoff to 48.3% at the second, culminating in an 89.3% escape rate at the third transition. By the final output, 23.7% of all injected hallucinations survive completely undetected in the finished executive report.
The Structural Ceiling of Post-Hoc Detection
Throwing larger reasoning models at terminal outputs cannot recover lost provenance. For GPT-4o, hallucination detection rates plummeted from 72.0% at Stage 1 down to 50.9% at Stage 4—a 21.1 percentage point decay. Even the strongest open architecture evaluated, Qwen3.5-397B-A17B, which flagged 87.0% of errors at the raw stage, hit a structural ceiling, with projected Stage 4 detection accuracy decaying toward 60% to 65%.
"When you verify matters more than whether you verify."
Because original context is laundered as raw data turns into fluent prose, end-of-pipeline verification fails by design. Singh and Pawar found that terminal evaluation yields a negligible 2.3 percentage point gain over doing no validation at all, leaving error survival at a dismal 58.4% (versus 60.7% unverified). In contrast, placing deterministic boundary gates and RAG verification directly at inter-agent handoffs slashes error survival to 16.2%—a 42.2 percentage point improvement over terminal checking (Cohen’s h = -0.911, p < 0.000001).
What this means:
For enterprise architects, the takeaway is unequivocal: multi-agent reliability is a problem of boundary-level propagation control, not downstream model intelligence. Relying on downstream reasoning LLMs or terminal prompt guardrails to catch upstream data corruption is architectural malpractice once numerical assertions have dissolved into qualitative summaries. Engineering teams must abandon passive prompt concatenation and implement strict, deterministic schema validation and source-grounded boundary gates at the earliest handoffs, where 75.4% of errors remain detectable, rather than burning inference budgets on futile post-hoc reviews.