I build private agent systems. The model is the easy part; my work is the harness around it: memory that survives the session, retrieval that cites its sources, and governance you can actually inspect. Local-first where the data is personal, cloud where it isn't.
The book, Private Agent Systems, is out on Amazon. It treats agents as production software, not prompt experiments. The code is here: Relic models who you are over time, so a reflective agent can carry your context forward without pretending to be you, while Amber answers over large document collections by fusing vector search with knowledge-graph reasoning.
