I wanted a Jarvis-style butler to run my life. After trying nearly every model as its 'brain,' I realized today's agents aren't omniscient assistants — they're routers that solve specific problems in specific scenarios. This is the three-tier setup I actually run, and what it taught me.
I came back to China and spent a stretch studying AI, with a pile of real personal needs I wanted to solve with it: weight control, health, job applications, learning, resume work. My old workflow — Notion as the main system, TickTick for todos, Flomo for fragments — was all manual upkeep. When agents arrived I tried to build a Jarvis-style butler with OpenClaw, tried nearly every model as its brain, and learned that today's agents can't be omniscient. The real lesson: an agent is best understood as a router that solves a specific problem inside a specific scenario, given a brain, a scenario-specific knowledge base, and an SOP. This is the three-tier setup I actually run today.
The Chinese version below walks through my original manual workflow, the attempt to build a Jarvis (and why it hit a wall), what I now think an agent really is — a scenario-bound router — the three-tier setup I run today, and what I'm building next.