How I think about agent design
A capable agent with broad access is a demo. A narrow agent you can actually reason about is a production system. Here's why I keep splitting agents apart instead of making them smarter.
Blog
Why I built VectorStep
OpenClaw is a brilliant personal AI assistant. Using it for real work made it obvious why "brilliant personal assistant" and "safe to run against production" are two different design problems - and that the second one didn't have an open-source answer yet.