Archibald, the system that runs my admin
In April 2026 I started building an AI agent system to run my own finances, email and schedule. I had no engineering background.
I built it for two reasons. My own admin, from bills and statements to email and scheduling, was taking hours every week. And I wanted to understand AI agents by building one, not by reading about them. My own life was the test case.
Archibald runs only my own accounts. It is separate from Archie, the product we sell at Otian AI.
What it does
It works on a schedule, mostly without me. It pulls transactions from six bank and card feeds, reconciles them, checks statements and closes each month. It also sorts my email and keeps my calendar. That gives me back about five hours a week.
I built it in Claude Code and Codex, in Python. Its skills have automated tests that catch regressions, and guardrails limit what it can do without asking me first.
- Scheduled runs since June
- 100,000+
- Successful
- 99%
- Jobs live today
- 145
When “ok” wasn’t true
In October a morning job stopped running for two days. The nightly check that schedules it reported “ok” three nights in a row.
The check confirmed that the scheduling command had worked. It never confirmed that the job would run. The job had been switched off by hand, and the command will update a switched-off job without complaint.
I changed the check to confirm the job will actually run, and added an alert after the first failure. Twelve of the 145 live jobs now exist only to audit the system’s own accuracy.
It’s the same lesson as the $1.8M power bill. A report that says everything matches isn’t proof. Check the thing itself.
The chart counts every job ever added to Archibald’s schedule, including ones later switched off, from its own job database. The run count comes from each job’s run sequence; the success rate is measured on the runs the log still holds.
Otian AI, from formation to launch →Questions about Archibald? Email me at jlraney99@gmail.com.