An AI research agent that shows you where every claim came from
Summaries are easy and mostly useless. What a research team needs is evidence: the passage, the source, the date, and an honest note when the answer is not there. Our research agents monitor continuously and produce briefs you can hand to a client without checking every line.
Example agent run, run · market-monitor-agent. plan: 5 sources, 2 queries each. tool call: sources.fetch(since="2026-09-01"). result: 31 documents, 6 changed. approval: publish brief?. approved: by research@client.com. tool call: brief.publish(cited=true).
How this work happens today
Monitoring is somebody checking twelve tabs on a Monday
The sources that matter are known. Checking them is tedious, so it happens weekly at best, and the thing you needed to know about moved on a Wednesday.
The summary loses the evidence
Someone reads twenty documents and writes four paragraphs. The paragraphs are good. But when a colleague asks which filing the second claim came from, the answer takes twenty minutes to reconstruct, if it can be reconstructed at all.
Nobody records what was not found
A search that returns nothing is a finding, and it is almost never written down. Three weeks later the same question gets asked and the same empty search runs again.
What a research agent actually does
It reads on a schedule, keeps track of what changed, and writes with its sources attached rather than remembered.
Watch sources and detect real change
Not just whether a page moved, but whether anything meaningful in it did. Boilerplate updates and reordered navigation are filtered out, so an alert means something happened.
Extract evidence, not impressions
The agent pulls the specific passage that supports a claim, with the document, the section and the date. A brief built this way survives someone senior asking where a number came from.
Show the retrieval trace
You can see which chunks were retrieved for a question, how they scored, and which ones were used. When an answer looks wrong, you can tell within a minute whether the retrieval or the reasoning failed.
Say when the answer is not there
The agent is tuned to report absence rather than fill it. A brief that says the filing does not disclose this is far more valuable than a fluent paragraph that quietly invents a number.
Produce the brief in the shape your team already uses
Your template, your section order, your citation style. Research output that has to be reformatted before it can be circulated does not get circulated, and the reformatting is where the citations usually get dropped. Matching your existing format from the start removes both problems at once.
Publishing is the write path here
A research agent reads far more than it writes, so most of its work runs unattended. The moment it publishes a brief, updates a shared knowledge base or sends a client facing document, that is an irreversible action and it stops for a person. What you approve is the finished brief with its citations visible, which is also the fastest possible way to review it.
How approval gates workWhat it connects to
Connectors ship behind feature flags, staged and reversible. Everything runs end to end in demo mode before a single live key is issued.
- Public web sources
- RSS and Atom feeds
- PDF and document stores
- Notion
- Confluence
- Google Drive
- PostgreSQL
- pgvector
Where to go next
Questions about this
How do you stop it making things up?
Two ways. Claims are generated from retrieved passages rather than from the model’s general knowledge, and the retrieval trace stays visible so an unsupported claim is obvious. We also test explicitly for it, using questions whose answers are deliberately absent from the corpus.
Can it work over our internal documents?
What about paywalled or licensed sources?
How current is the output?
Give us five sources you check by hand
We will show you what continuous monitoring over them looks like, with the citations attached.