Designing agents to manage knowledge, not just generate text
Abstracts
Abstracts
A recurring focus in this work is agent product design applied to knowledge, not just output. The crawler agent prototype and ProfileScribe both center on the same problem: scattered professional evidence needs to be normalized into something a system can reason about before any user-facing content gets written.
That framing changes what counts as progress. It's less about how fluent a draft sounds and more about whether the underlying facts it's built on can be traced, checked, and trusted. That's the part of agent design that's easy to skip and hard to retrofit later.