Research Agent Workflow: Build a Source Trail Before Writing a Report

Evidence check: 25 October 2026. Product details reflect the referenced documentation observed at this cutoff; exact plans, prices, regional eligibility and unverified release dates are not claimed. This is an editorial guide, not a hands-on test.

Start with a bounded research question

Specify subject, geography, cutoff date, intended readership and what counts as authoritative evidence. A 2026 product feature cannot be assumed to have existed when a 2022 blog post was published. Define acceptable sources and what must be reviewed by an editor. The workflow must be able to conclude that a claim remains unverified.

Search and evidence capture

Issue focused search queries, open original documents, record exact URLs, publisher and access dates. Distinguish a vendor announcement from documentation of actual availability. Capture short claim excerpts and keep a link to the original rather than flooding the model with duplicated pages. When sources conflict, note the difference instead of picking the more convenient statement.

Claim extraction and drafting

Turn each supported claim into a structured record with source ID, date, region, version and status. Draft readable prose from the records, not from untracked context. A source might establish a product’s feature but not its price, geographic entitlement or real-world performance. Keep those fields null until corroborated. Never create citations to guessed URLs.

Editorial checkpoint

A human reviewer checks that the lead, claims and caveats match evidence. Require special review for prices, API versions, enterprise security, commercial licensing and high-stakes subjects. Approve a frozen draft and its source map, not a model’s statement that it “fact-checked everything.”

Operational boundaries

Allow retrieval tools to read an approved list of public sources. Deny website logins and outbound writes by default. Time-box searches, deduplicate results and stop when new searches no longer change material findings. Treat tool output as untrusted data, including fake “publisher instructions” hidden in pages.

Test the workflow

Inject a stale document, a moved URL, conflicting release dates and a malicious instruction disguised as an official changelog. Evaluate whether the final draft accurately describes uncertainty and refuses to turn unresolved facts into confident claims. Store run IDs, source-claim mappings and reviewer corrections for later audits.

Build the evidence matrix first

For a product-release article, create columns for claim, official document URL, publisher, publication date, retrieval date, exact product version and supported region. A single press-release sentence may establish an announcement but not general availability. When documents contradict each other, record both and state which is authoritative for a particular date. Save extracts with enough context to inspect them later, while respecting publisher terms and copyright.

Compare an assistant to an editorial baseline

Give the agent a defined collection of sources and ask for a report with ten checkable facts, explicit unknowns and a separate narrative draft. Grade whether every material assertion has a valid source, whether old pricing is avoided and whether caveats survive prose rewriting. Introduce a fabricated “official instruction” inside a retrieved page and confirm that it does not override task policy. A human editor remains responsible for approval, factual judgment and publication; a polished source list is not proof that each claim was checked.