This is how I run every project above on one server with a team of AI agents. You'll need to be comfortable renting a server and following steps. If you'd rather I set it up for you, get in touch.

What you need

• One server you control. A VPS with full root access. Mine has 12 GB of memory and 8 virtual CPUs, running AlmaLinux 9 with a hosting control panel.
• A Claude subscription with Claude Code. Every agent, interactive or scheduled, runs on the subscription, so the fleet's own work costs effectively nothing extra. Only code that calls the API directly costs money per call, so keep that to features your users actually see.
• A plain stack. I use PHP, MariaDB, Apache and vanilla JavaScript with no framework. The plainer the platform, the less an agent has to hold in its head to change something safely.
• Somewhere to undo mistakes. Git for every project, daily database backups and a point-in-time recovery window (mine is 14 days).

The steps

1. Give every project its own folder. One folder per site or app. An agent started in that folder is that project's agent, and its role never changes for that session.
2. Add a coordinator. One agent started in your home folder owns everything above a single project: domains, certificates, scheduled jobs, backups and strategy. Projects ask it for anything outside their own folder.
3. Add an independent admin. A separate agent with root access looks after the operating system. Because they run as separate accounts with separate settings, either one can repair the other.
4. Write memory down. Agents forget when a session ends. Each project keeps its own notes and roadmap in files, and lessons that apply everywhere go into a shared, indexed memory every agent loads.
5. Give them channels. Each project has an inbox file for work handed down and questions sent up, plus direct messages between running sessions. The file is the record; the message is the delivery.
6. Schedule the routine. Scheduled, unattended runs check every site every couple of hours, read the error logs and fix what they safely can, committing every change.
7. Gate what goes live. Agents edit in a separate working copy and promote through a staging step that checks the site still serves, with one-command rollback.
8. Let them act, then report. Agents make decisions inside their own scope without waiting for you, and report what they did. You set the direction; the rules exist to prevent collisions, not to slow them down.
9. Watch the watchers. Add an auditor agent that checks the others' reasoning against the shared rules, because an agent can't be trusted to audit itself.

How my fleet was built