How I think about agent design
[PLACEHOLDER — draft skeleton, not full prose yet. Sketch below is meant to be expanded, not published as-is.]
The instinct when building an agent is to make it more capable: give it more tools, let it handle more of the job, trust it to figure out the rest. That instinct is right for a personal assistant — it’s basically what makes OpenClaw compelling. It’s wrong for anything you’d call production.
The temptation
Section titled “The temptation”- One agent that triages, investigates, and remediates feels efficient. It isn’t — its confidence number ends up meaning three different things depending on which part of the job it was doing at the time.
- “Just give it all the tools, let it decide” feels flexible. It’s actually the opposite: more tools is a bigger blast radius and a noisier decision space, not more capability where it counts.
What I do instead
Section titled “What I do instead”- Split by responsibility, not by convenience. Each agent earns a narrow job description and a confidence number that means one specific thing.
- Tools are a capability grant, not a courtesy — an agent can’t reach past
its
tools:list, so keeping it short is a real security boundary, not just tidiness. - Confidence has to measure “did I do the job,” never “how scary does this look” — the two get conflated constantly, and it quietly wrecks calibration.
The practical version
Section titled “The practical version”The engineering checklist version of this post lives at Writing good agents — this post is the “why,” that page is the “how.”