Key technical lesson from LinkedIn's agent build-out (see Narrow, Single-Purpose Internal Agent Suite (LinkedIn)): giving an agent blanket access to a shared drive or general knowledge base performs poorly and causes hallucination. The fix is curating a small, high-quality "golden examples" set for the agent to draw on instead of dumping the entire corpus.
Tomer Cohen traces this lesson back over a decade, to manually curating a set of "good LinkedIn post" examples while rebuilding the feed team — the same curation-over-volume principle, applied first to a ranking/recommendation problem and later to agent grounding.
Tomer Cohen, LinkedIn CPO, on why golden-example curation beats blanket access: 'it's not great to just give it access to your drive and say reason all over this knowledge base' — and on what fixed it: 'the first and most important part was fitting in the right data, not all the data.'
Cohen's blunt evidence for curation over blanket access: pointing an agent at an entire knowledge base or shared drive 'failed miserably and hallucinates like crazy' — blanket access causes the agent to weight unimportant material as if it were load-bearing. The fix wasn't a smarter model, it was stricter curation: a hand-picked 'golden' example set. He frames this as the single most effective agent-building technique LinkedIn found, ahead of any model upgrade.