Case study: Cultural heritage research

Multi-Agent Reputation and Discovery Pipeline

The same archive already has a grounded chat agent, Victor's Docent, answering visitor questions at victorstember.com. This is a different system: a multi-agent research pipeline that swept the open internet for the family's own catalogue, for works outside their collection, and for misattributions no one had caught, across five rounds against 238 known works.

Visit victorstember.com

The challenge

The family's 238-work catalogue had no way to learn what auction houses, museums, and foreign-language sites already held or had sold outside the family's own records
Manually cross-checking hundreds of candidate matches against a 238-work catalogue, across languages, does not scale to one researcher's hours
A wrong match recorded as fact would corrupt scholarship the family had spent decades building
Provenance leads and public authority records sat unlinked, leaving the catalogue with no verifiable trail into the wider public record

What we built

A multi-agent research workflow, not a chatbot: independent sweep lanes worked the open web in parallel, each reporting candidate findings back for cross-checking against the existing catalogue before anything was recorded as fact.

37 sweep lanes ran across 5 rounds, followed by a dedicated 10-lane round built specifically to close remaining completeness gaps
36 of the family's 238 catalog works were independently rediscovered on outside sites, including 7 works the family holds only as a thumbnail
32 candidate new works were reviewed, and 2 were confirmed as genuinely new to the catalogue
2 existing records were corrected after verification: art-197 retitled Portrait of Grand Duchess Elizaveta Feodorovna, previously misattributed to Empress Alexandra Feodorovna, and art-210 retitled Portrait of a Young Nun after direct visual comparison to a Lipetsk Museum holding
A Wikidata item and three Wikipedia articles, in Russian, Italian, and Japanese, were added as authority links, giving the catalogue a verifiable trail into the wider public record

How it's built

A multi-agent workflow orchestration, separate from the chat agent that answers visitor questions on the same archive: independent subagents ran research sweeps, reported findings for cross-checking against the existing 238-work catalogue, and escalated only verified matches for a recorded human decision. The run spanned several resumed sessions from September 4 to September 8, 2026, roughly 800 to 900 subagent calls in total, entirely on the Claude subscription channel rather than a metered API key, so the marginal cost of the whole sweep was zero.

Results

36 of 238 catalog works independently rediscovered outside the family's own records, seven of them previously known only from a thumbnail
2 new works confirmed and 2 misattributions corrected, each backed by recorded verification rather than a guess
Wikidata and three-language Wikipedia authority links now give the catalogue a verifiable trail into the public record
The entire five-round sweep, about 800 to 900 subagent calls, ran on the Claude subscription channel at zero marginal cost

What is already being said about your work, somewhere you are not looking?

A research pipeline like this can run against a catalogue, a reputation, or a body of work nobody has time to search by hand. We build it as a project, scoped to what you actually need found.