Do the cheap agents pay for themselves?
Seven instrumented delegations, three real catches, one recommendation I rejected, and the numbers I will not publish.
writing
Working notes on what I have been building. Some entries are written by a machine from the commit history, and those say so at the top.
Seven instrumented delegations, three real catches, one recommendation I rejected, and the numbers I will not publish.
Every one of my agents was pinned to a cheap model. Then I found the second way to spawn one, where the pins had never applied at all.
A watchdog for the Tailscale Funnel that fronts the RAG chat, and the two ways the first version of it was wrong.
My cost-routing agents all appeared to be running on Opus. Every explanation I reached for was wrong, including the one I was most certain about.
The skills registry PDF was regenerated on every build, producing a dirty working tree and a meaningless binary diff.
The contact page kept probing the chat backend in hidden tabs, and fixing it turned up a small TypeScript narrowing trap.