Designing agent products around what evidence actually supports
Abstracts
Abstracts
Working on ProfileScribe and its crawler prototype together has clarified a general pattern in agent product design: the hard part isn't getting an agent to produce output, it's constraining what it's allowed to claim. Both pieces exist to normalize evidence before it reaches a person's professional memory, and to keep drafting bounded by what a source actually shows.
That constraint-first approach is less flashy than open-ended generation, but it's what makes agent-drafted content something a professional could stand behind. The interesting design work is in the guardrails, not just the output.