Inside the system
How we run this business on Claude.
Every consultant says AI changes how you work. This is ours, running now: thirty-six agents on a schedule, a single source of truth they all read from, and a set of rules that took a year of breaking things to write down.
We teach this because we run on it. Not the other way round. And the same hands that built it build for clients: the workflows below have counterparts running inside client environments, built in their accounts during Rollouts and build sessions.
What runs, and when
No servers. Scheduled jobs, every one of them also runnable by hand.
The architecture
One principle carries all of it: a sensor may not write, a writer may not decide, and every change is preceded by a proposal that carries its evidence. Most automation fails because one script both decides and writes, and nobody can say afterwards why a field changed.
1
Sense
Read-only observers turn real events into signals. An email that sends a proposal, a meeting with action items, a payment landing. Nothing here can write.
2
Propose
Every intended change becomes a record first, carrying its evidence: the message, the charge, the transcript that justifies it. No proposal, no write.
3
Arbitrate
One component owns the queue. It removes duplicates, resolves two agents wanting the same thing, and decides what is safe to apply without asking.
4
Confirm
Applied automatically only when the evidence is a hard fact and the change is reversible. Everything else is one tap in Slack.
5
Write
One owner per field. Not one owner per system - per field. Two writers on the same field means the last one to run wins, which is not a system.
6
Verify
A write that was not checked did not happen. Every job reports how many records it changed, and a separate check asks whether the source of truth still matches reality.
What it is designed against
Every rule here exists because something broke first. These are the ones worth stealing.
A green job is not proof anything happened
A scheduled job that exits cleanly having done nothing looks identical to one doing its work. We had a data sync report success every day for eleven weeks while writing to a database nobody read from. Every check now reports how many items it examined, and examining zero fails loudly instead of passing quietly.
A number with two homes will disagree with itself
The same figure stored in two places diverges silently, and you find out when two reports contradict each other in front of a client. Every number now names the one store that owns it.
Never sum a ledger without filtering it
Transfers between accounts cancel real activity. A total that does not exclude them is confidently, catastrophically wrong in whichever direction you were not expecting.
A monitor that cries wolf gets muted
Our health monitor marked itself failed whenever it found anything, so it was red for twenty runs straight. When a real outage arrived, it was invisible inside its own noise. Red now means the check broke. Findings go somewhere a person is looking.
Want this in your business?
The barrier is knowing what to build, not how to code it. That is what we teach, and what we build with teams who would rather have it running than have it explained. In a Rollout your highest-ROI workflows get built into your own environment. In the Build Series we ship one working thing per session with your team in the room.
Or read how a single Claude Project becomes a chief of staff - the smallest version of the same idea.