How Vivan runs on AI agents
Our own company, run by departments of AI agents, with a person approving what matters
Vivan runs its own work the way we set it up for clients: departments of AI agents for the company site, each product and marketing. A manager agent writes the tasks, builder agents do them on a separate copy, the manager reviews the work and sends it back, and anything on the red list waits for the managing director's tick in Taskinger.
One department, one week (example)
- 14 Tasks written
- 5 Sent back with fixes
- 11 Released and checked live
- 2 Waiting for your tick
- 7 Stopped by the guard
Who is it for?
This is for a business owner who wants to see the AI agent departments service at work before paying for it. The setup on this page is the one we use every day for our own sites and products: SnapReceipt, Taskinger and Mamino are built, reviewed and released through it.
What it does
A department per area
One department each for the company site, SnapReceipt, Taskinger, Mamino and marketing. Its manager agent owns every task from the request to the check on the live site.
Builders that cannot release
Builder agents work on their own branch of the code, run the checks and report, and a second model reviews the finished branch. They cannot merge or release anything.
Review and send back
The manager reads the whole change, the screenshots and the checks, and sends it back with numbered fixes until it is right.
A red list for the important things
Money, prices, clients, legal text, security rules and store releases wait for the managing director's tick in Taskinger. Everything else the manager releases itself and reports.
A guard before every action
Every action of an agent passes a guard first. It stops sending mail, paying, deleting files, opening keys and releasing any other way before the action, not after it.
Releases with a rollback
One script does every release: checks in a clean copy, merge, deploy, checks on the live site, and an automatic rollback when one fails.
A supervisor over the managers
A separate supervisor agent checks the departments' work against its sources three times a day, including the managers' own reviews.
Built with
Claude runs the managers, Codex and Claude do the building, Codex reviews finished work, and Gemini (Antigravity) takes part of the marketing. Tasks and approvals live in Taskinger, where the agents are workbook members. The scripts around them are plain Python with no AI in them, and every run leaves one line in a log.
Where it stands
In daily use since October 2026 for our own sites and products. The figures in the picture are examples for illustration, not our real numbers: we do not publish internal figures.