44 sites, one engineer, a fleet of agents

the problem

Running dozens of production sites alone means the operations work scales with the number of sites unless something else does it. Handing that work to AI agents is only acceptable if they are cheap when idle, bounded when busy, and unable to skip the guardrails.

Constraints

• Most ticks find nothing wrong. Those ticks should cost nothing.
• Agents get budgets: turn caps and wall-clock limits per role.
• Nothing deploys from a laptop. Every change ships through GitHub and Cloudflare Workers Builds.
• Anything written in my voice is a draft until I read it.

Stack

how it is built

1. 01

   Scheduler

   A cron container per site. A watchdog every 15 minutes, an engineer every 30, an error scan every 5, and weekly SEO, planning and writing roles.

2. 02

   Bash first

   Each role starts as a shell orchestrator doing the mechanical checks with zero model calls. The model pass only runs when the sweep finds real work.

3. 03

   Gates

   Commits go through git hooks and a production build. No bypass flags, no override variables.

4. 04

   Ship

   Push to GitHub; Cloudflare Workers Builds builds and deploys. One Worker per domain.

5. 05

   Observe

   A custom Cloudflare API telemetry collector (cf-stats) feeding Grafana dashboards, a FastAPI/SQLite site tracker, and uptime checks that alert to Slack. Every Slack alert is also written to disk, where the error scan picks it up.