The Annual Pass That Cost More on Paper

In a multi-agent financial-planning task, a "Tickets_Pricing_Expert" agent was responsible for gathering ticket pricing information and produced incorrect figures, then stopped short without completing the comparison and summary steps it was meant to hand off. A "Budgeting_Expert" agent downstream took the flawed pricing data at face value, used it to calculate costs, and concluded that an annual pass was more expensive than a pay-per-visit option, the opposite of the correct answer. A separate "Verification_Expert" further downstream, working only from the already-corrupted figures it had been handed, had no way to catch the original error.

The Verification Expert failed at its job, considered and ruled out; this is in fact the mistake other automated failure-analysis approaches made in this same case, blaming the downstream agent for the wrong output when it never had access to correct figures to verify against in the first place.

Apraxia on the part of the Budgeting Expert, considered, since it produced a wrong calculation. Ruled out because its arithmetic on the numbers it was given appears to have been correct; the input it was reasoning from was already false.

Shared Psychotic Disorder, most consistent with the evidence. An early, incomplete error from one agent propagated unexamined through two further agents in the chain, each of which trusted the information it inherited rather than independently confirming it, so the final wrong conclusion looked, at every later step, like a reasonable one.

This is the same reason a mistaken assumption spreads through a human team when no one asks "where did that number actually come from", except a chain of specialized agents can pass information along even faster and with even less friction than a group of colleagues would. Multi-agent pipelines built around passing a running belief from one specialist to the next need at least one point where a number or claim gets checked against its original source, not just against whether it's internally consistent with what the previous agent said.

Add a point in the chain where a number or claim is checked against its original source.

Source verification between agents; provenance tracking of figures; independent checking beyond internal consistency.