“If your agent has been running for a week and you cannot say what it did on Tuesday, you do not have an agent.”
The Journal · Vol. I · Issue 06
Field notes for people shipping real agents.
Notes on the practice of building agent systems that can be inspected, paused, handed off, and improved without turning every workflow into a mystery box.
Field notes · 12 min read
Why every agent should have a HEARTBEAT.md
Monitoring is not a cute ping. It is how you notice stale sessions, quiet failures, and unresolved owner decisions before they become a broken system.
Read in full02Safety · 8 min read
Stop conditions
The fastest way to make an agent trustworthy is to name the work it must refuse before you name the work it should do.
03Build-log · 6 min read
Friday-morning runbook
A weekly operating ritual for reviewing what agents touched, what they skipped, and what needs owner attention.
04Design · 9 min read
What we got wrong building v1 Builder
The first Builder tried to be clever. The better version asks four boring questions and gives you files you can edit.
05Voice · 5 min read
Voice without sounding like chatbot
Reusable agents need house style, but they also need a line between helpful tone and fake personality.
06Changelog · 4 min read
Spring 2026 update
What changed in Vol. I: cleaner stop rules, sharper SMB examples, and a Builder that exports plain Markdown.
07Team · 7 min read
Why we name agents after weather systems
Names are operational handles. A good one lets a team talk about ownership, status, and incidents without confusion.