The job
“Give me a defensible picture of this market by Thursday — and I need to know which numbers I can put in front of the board.”
Desk research is not slow because reading is slow. It is slow because checking is slow. Anyone can assemble twenty links in an afternoon; confirming that each one actually supports the sentence it is attached to is the part that eats the week — and it is exactly the part that gets skipped under deadline.
That is also where language models are least trustworthy. A model will produce a citation-shaped string with total confidence. If nothing opens the link, nobody finds out until a reader does.
What Forge does
Two agents with different jobs, and the second one does not take the first one’s word for anything.
The researcher works the question — searching, reading rendered pages (not just raw HTML), following what looks promising, discarding what doesn’t.
The analyst then re-opens every source independently and checks that the cited claim is really there. Three outcomes, and all three end up in the report:
- Verified — the analyst retrieved the page and confirmed the quote.
- Corroborated — the original was unreachable (bot protection, paywall), but the fact was confirmed from another source, which is named.
- Unverifiable — nobody could confirm it. It stays in the report labelled as unverifiable, or it comes out.
That third category is the one that matters. It is the difference between a research tool and a plausible-text generator.
Evidence
A real run on the Russian market for business-process automation tooling, commissioned as a live test:
- 18 sources gathered by the researcher across news agencies, industry press, corporate publications and technical media.
- 11 verified by direct retrieval — the analyst fetched the page and confirmed the quote itself.
- 4 blocked by bot-protection; each fact re-confirmed from a named alternative source.
- 2 figures marked unverifiable in the final document rather than presented as fact.
- Output: an 8-page PDF, working links throughout, correct Cyrillic typography.
Eleven of eighteen checked by live retrieval, the rest disclosed. No invented citations in the deliverable.
What you control
- Depth and budget. Research runs consume tokens; you cap the spend per run, and the run stops rather than quietly escalating.
- Sources. Point it at the open web, at your own systems through connected tools, or both.
- The gate before delivery. The analyst’s verdict is visible before anything reaches you — including when the verdict is “this question could not be answered well”.
Under the hood
- Workflow Engine — the two-role chain and the validator between them
- Multi-Provider Routing — which model serves which step, and what it costs
- Git-Backed Audit — every intermediate handoff kept, so you can see how a conclusion was reached
- Observability — why a given step behaved the way it did