Fuselab: agent interfaces need a separate activity layer, not a chat bolt-on

Fuselab's argument is that agent interfaces need an entirely different information architecture than chatbots, not incremental additions to one. Drawing on regulated-industry work like a Medi-Cal enrollment system for DHCS, the agency insists every automated decision must stay traceable to the caseworker responsible for it. The piece names four requirements, transparency, override, proactive status, and structured recovery, and backs the trust point with NN/g's finding that users rarely click the citations they claim increase their confidence. The practical takeaway is architectural: separate the conversation thread from the activity log, because chat cadence and workflow cadence genuinely conflict.