Build Research Agents That Preserve Their Evidence

A research agent should preserve where each claim came from, when it was observed and what remains uncertain as findings become summaries.

A research agent can return a polished answer that is difficult to audit. The source list may look credible, yet the reader cannot tell which page supports which claim. When the information changes, the next run has no reliable way to explain the difference.

I prefer to treat research as a sequence of evidence transformations. A source is retrieved, a relevant observation is extracted, a claim is formed and a conclusion is written. Each transition should retain enough context to trace the result backward.

Make claims smaller than documents

A source page can support one fact while remaining silent about another. A company website might describe installation services without confirming every region where they are offered. Storing only the page URL makes it tempting to attach that source to a broader claim than it supports.

I would retain a claim record with its source location, observation time and relevant supporting passage or field. The record should also identify whether the claim is a direct statement, an inference or an unresolved possibility.

Building PVFirms, my solar-industry directory, provides a useful context for this distinction. A directory needs coherent company information, and a well-written description should not silently fill gaps in the underlying facts.

Preserve what the source does not establish

Missing information is an ordinary research result. If a service area is not stated, the agent should record that limitation. Nearby offices, familiar terminology or an attractive map do not automatically establish service coverage.

When sources disagree, the system should preserve the conflict before deciding how to present it. It may prefer a newer first-party statement, seek another source or mark the field for review. Merging two incompatible claims into one smooth sentence hides the decision.

The same applies to dates. A page’s retrieval time is not necessarily the time the information became true. I would keep observed, published and effective dates distinct when the source provides them.

Separate extraction from editorial writing

Extraction asks what the material supports. Editorial writing asks how to express the accepted findings clearly. If both happen in one unconstrained step, the desire for a complete paragraph can encourage unsupported additions.

A useful interface lets the writing stage see accepted claims and explicit gaps. It can improve readability without changing a company’s identity, inventing a certification or converting a tentative statement into a guarantee.

This also reduces the risk of treating source instructions as authority. The extracted evidence should carry its origin, while the application retains control over the publication workflow.

Check the claim-to-source relationship

A citation is helpful only when it supports the sentence beside it. I would review a sample of claims by reopening the source and asking whether an independent reader could reach the same limited conclusion.

For facts that change frequently, the review also needs a freshness policy. Reusing last month’s successful extraction may save work, but it does not make the information current. The system needs a reason to refresh, retain or flag the record.

Source access can fail as well. An unavailable page should produce an unavailable-source state, rather than being treated as evidence that the company or service no longer exists.

Let the job finish with uncertainty

A research task can be complete even when some requested fields remain unknown. Completion means that the agreed research process reached an acceptable result, including its limitations. It does not require every field to contain a confident answer.

Explicit stopping conditions help prevent repeated searching from becoming a substitute for acknowledging a gap. The final output should make supported facts easy to use and unresolved questions easy to identify.

Preserved evidence turns research into maintainable data. It lets the next person update, challenge or reuse a finding without reconstructing the entire investigation from a polished paragraph.

Updated 30 September 2026.