We need to tap into the institutional knowledge to make AI agents effective. But the most valuable knowledge is tacit: trapped in people's heads, pre-articulation, never written down. How do you surface it?