Historical Information Discovery: When AI Agents Work as Researchers With a Hundred Eyes
The question: what actually happened? When the information about a historical event is partial, contradictory, or simply missing from the places you'd expect to find it — how do you assemble the puzzle?
In this post we share a project that faced exactly this challenge: reconstructing an only-partially-documented event, using AI research agents working in parallel across dozens of sources.
The Background: A Half-Erased Event
Our client — a research institute — had a specific question: reconstruct the details of a specific event from the 1950s. Official documentation of the event was sparse, contemporary press yielded only brief and vague mentions, and there was reason to suspect that some documents had been deliberately removed from the archive.
The decision: not to rely on "visible" sources alone, but to cast a wide net across every edge of accessible information — hoping that indirect evidence might fill the gap.
The Classical Approach (And Its Failures)
In the past, this kind of research would have required:
- Reading thousands of pages of minutes
- Physically searching in more than 20 different archives
- Interviews with the few survivors who might remember
- Locating descendants of secondary figures who might have kept private papers
- Cross-referencing dates and details across disparate sources
Estimated time: 5-7 years. Cost: very high. Chance of completeness: low.
The New Approach: A Network of Research Agents
We built a team of specialized LLM agents, each with its own "role", working in parallel:
Agent 1: Contemporary Press Researcher
Read a focused sweep of every issue of 14 major newspapers from the two years before and after the event, searching for:
- Direct mentions
- Indirect mentions (via names of relevant people and places)
- Advertisements, announcements, obituaries
- Reports framed as "other events" that might be the same incident in disguise
Agent 2: Government Records Researcher
Worked in parallel on relevant minutes published on government websites and state archives, identifying:
- Mentions of the names of involved figures
- Debates in the relevant periods
- Decisions that look like "responses" to the event, even without mentioning it directly
Agent 3: Memoirs and Biographies Researcher
Digitally scanned over 200 biographies, autobiographies, and letter collections of period figures who could have been involved or informed — searching for mentions, hints, or "loud silences."
Agent 4: Cross-Reference Researcher
Received every finding from the other agents and tested it against:
- People databases (Who's Who, Judaica databases)
- Event databases (chronicles, subject indexes)
- The calendar — checking whether specific details "ring together" consistently
Agent 5: Report Writer
Aggregated all findings into a structured document, with full citation for every source, a confidence level for every fact, and recommendations for research directions the data pointed toward.
What We Found
This research revealed a picture far more complex than the contemporary press suggested:
- 7 additional figures who had been involved, whose names didn't appear in any public source
- A detailed timeline of the 11-day event — which contradicted the official record on 3 key days
- At least 4 documents located in private family archives — whose existence no one knew of
- A network of relationships between figures previously thought to be disconnected
The Most Important Lesson
The lesson we took from this project: not all research requires the same kind of AI.
- For text search — RAG
- For pattern recognition — LLMs with focused prompts
- For multi-source investigation — autonomous agents with defined roles
The orchestration of these tools — which agent passes which information to which other, who verifies, who summarizes — is our professional added value.
Time and Outcome
Total time: 4 months. Cost: less than 15% of the traditional approach estimate. Quality: a 180-page research document, with 600+ footnotes and references, that enabled the senior researcher to write an academic paper published a year later.
Your Hidden Research
If you have a historical question you feel there's no way to find an answer to — perhaps there is. The right combination of technology, archival methodology and patience can reveal what was thought to be erased for generations.
