A useful research agent turns a scoped question into subquestions, searches approved sources, records source identity and exact evidence, compares conflicts, and writes only claims supported by that record. It should distinguish quoted fact, synthesis, and open question. Source tracking is data captured during research, not a citation list invented after drafting.
Who this is for: Researchers, analysts, and developers building assistants that gather and synthesize evidence from approved information sources.
- Define the research question, source boundaries, freshness need, and completion rule before searching.
- Store evidence at claim level with location, access date, source type, and support or contradiction status.
- Require explicit gaps and conflicts so the final report does not hide weak coverage behind fluent synthesis.
Scope the question and evidence standard
Turn a broad topic into a decision or answer with named dimensions. 'Research electric fleets' is unbounded. 'Compare operational requirements for replacing twenty local delivery vans, excluding vehicle purchasing advice' gives the agent a purpose and limit. State geography, time period, acceptable source types, and which claims require primary or official material.
Define completion as evidence coverage, not a fixed number of searches. A report might need a supported answer for charging, route range, facility constraints, and regulatory obligations, plus a list of unresolved cost inputs. Set time and search limits, but let the workflow stop early when all criteria are met or clearly blocked.
Capture sources during discovery
Every result should enter a source ledger with title, organization or author, durable URL, publication or update date when available, access date, source type, and the exact passage or data supporting a candidate claim. Preserve the original wording and location. A URL alone forces later reviewers to repeat the search and may not show what the agent relied on.
Treat retrieved pages as untrusted data. Text can contain instructions aimed at an automated reader. The research tool should extract content without allowing page text to change system rules, tool permissions, or report destinations. Apply domain and access policies in code. Do not follow requests inside source content to reveal secrets or contact third parties.
Build a claim-evidence map
Before drafting prose, create proposed claims and link each to supporting and contradicting evidence. Label direct observations, calculations, and synthesis separately. A source may establish a regulation's text but not its effect on a particular company. The agent should narrow the claim or mark the application question for qualified review.
Compare dates, jurisdiction, definitions, and source authority when material conflicts. Do not resolve disagreement by counting links. Prefer the source appropriate to the claim and explain remaining uncertainty. Duplicate articles that repeat one announcement are not independent confirmation. The ledger can group derived coverage under the originating primary source.
Generate and audit the report
Draft from the claim map, attaching citations through source identifiers controlled by the application. Require every factual assertion that matters to point to evidence. Include a methods note covering scope, search dates, exclusions, and limitations. The conclusion should distinguish findings from recommendations and name assumptions a decision-maker must verify.
Evaluate research with known questions and source packs. Score coverage, claim support, citation correctness, source appropriateness, conflict handling, and unsupported additions. Test an unanswerable question and a malicious page. Search and tool APIs change, so check current official documentation quarterly and revalidate integrations instead of assuming old browsing behavior remains available.
Research a software accessibility requirement
A product team asks what an upcoming procurement requirement means for a planned release in one jurisdiction.
- Clarify jurisdiction, effective date, product type, procurement context, and the boundary between factual research and legal interpretation.
- Split the work into requirement text, applicability, dates, technical references, and unresolved product facts.
- Search official sources first, saving exact passages, dates, locations, and access details into a source ledger.
- Map each proposed statement to evidence, flag conflicts or missing applicability facts, and route legal conclusions to counsel.
- Produce a report with claim-level citations, methods, gaps, and concrete questions the product and legal teams must answer.
Claim and source ledger
Use one row per material claim so support can be reviewed before prose is polished.
- Claim: precise statement, scope, date sensitivity, and whether it is fact, calculation, or synthesis.
- Evidence: exact passage or data, source identifier, location, publication date, and access date.
- Assessment: supports, contradicts, partially supports, or does not establish the claim.
- Authority: primary, official, secondary, or anecdotal source plus relevance to jurisdiction and period.
- Disposition: include, narrow, omit, search further, calculate, or send to qualified review.
Common mistakes
- Collecting links first and asking the model to invent which source supports each sentence after the report is written.
- Treating several articles derived from one announcement as independent evidence for a claim.
- Allowing instructions embedded in a web page to alter search scope, tool access, or destination of the research output.
Try one
A research agent finds two official pages with different deadlines. What should it do before drafting a conclusion?
It should record both exact statements, publication or update dates, jurisdiction, audience, and document authority. It should look for supersession notices or a governing primary document, not pick the newer-looking page automatically. If the conflict remains, the report must state it and identify the owner who can resolve it. A passing answer preserves both citations and narrows any claim to what the evidence actually establishes.
Sources
- OpenAI agents guideOfficial guidance for building agent workflows with current OpenAI platform capabilities.
- OWASP Top 10 for Large Language Model ApplicationsOfficial OWASP project covering major security risks in applications that use language models.