BLOG
Notes from building AI agent systems and automations — what breaks, what actually works, and what running a dozen parallel AI sessions at once teaches you about software that runs itself.
Today one AI apologized to another
What it looks like when one agent's fix breaks another agent's assumption, and they have to work it out.
Rebuilding the app while it's running
Shipping a redesign to a live product without a maintenance window, and without users noticing the seams.
Making three AIs referee each other
Why one model reviewing its own work is worthless, and what changes when you make three of them argue.
The security fix that locked itself out
A deploy-role permission fix that, applied correctly, denied its own ability to ever fix itself again.
The worktree that looked empty
A destructive command run one directory off from where I thought I was, and the discipline that should have stopped it.
I run a dozen Claude sessions at once
How I actually coordinate a dozen parallel agent sessions without losing the thread — or my sanity.
Tests that a stub can pass
The predicate that stayed green while the feature underneath it was never actually built.