Agent Ops Starter Kit
A starter kit and field guide for running AI agents safely without supervision.
- Status
- v0.9
- Year
- 2026
- Role
- Solo
- Area
- Agents & AI systems
In plain English
Running AI agents unattended goes wrong in quiet, surprising ways. This kit is the operating manual I wish I'd had, distilled from incidents on my own agent fleet. Everything runs offline against test data, with no accounts, no keys and nothing that phones home.
How it works
- Schedule agent jobs with ready-made launchd templates (the Mac's built-in scheduler).
- Apply cost gates that stop a job before it overspends.
- Route work through a review gate before it counts as done.
- Follow the runbook "12 ways your agent fleet will betray you" when something looks off.
- Test everything offline with pytest against fixtures (prepared sample data).
What I did
I wrote it solo, based on what went wrong on my own fleet.
What I learned
Many failures look like success, so I check the actual output. A running process, a clean exit and a "sent" message can all hide a system doing nothing useful.