How I run a 33-agent autonomous business from one desk

A year ago CBX Nexuswas one eBay store, one spreadsheet, and a lot of manual listing work. Today it's a 33-agent fleet running across six marketplaces, coordinated by a Next.js dashboard and a Supabase-backed orchestrator. One operator. Same desk.

This is a rough map of how that works.

The fleet

Every agent is a specialist. There's an eBay sourcer, an Etsy designer, a KDP interior generator, a Fiverr responder, a GitHub opportunity scanner, a YouTube scripter, a YouTube renderer, a content pipeline that feeds the next one, and so on down the chain.

They're not all "AI" agents. Some are Python scripts running on cron. Some are Node workers listening on a queue. A handful are Claude-driven with tool access. The common contract is: every agent writes to event_log, pings health_checks, and reads its work queue from batch_jobs.

The orchestrator

The orchestrator is a boring Node service. It schedules jobs, fans out work, and respects per-agent rate limits (eBay's API budget is the tightest constraint in the whole system).

It doesn't try to be clever. Clever orchestrators are where autonomous systems go to die — you can't debug an opaque scheduler at 2am. Mine writes every decision to the event log with enough context to replay it.

The self-healer

The single most valuable piece of code in the stack. Every 5 minutes it scans health_checksfor anything that hasn't reported in a while, and batch_jobsfor stuck work. If an agent hasn't pinged in 15 minutes, it gets restarted. If a job has been in-progress for too long, it gets retried on a clean worker.

Before the self-healer existed, I'd wake up to a dead Fiverr responder and a pile of angry customers. Now I wake up to a green dashboard and a log of what the healer caught.

The dashboard

Built in the same Next.js + Supabase stack as this site. Two jobs:

• See the fleet. Every agent, every status, every last event — in one view.
• Override the fleet. One-click pause, restart, or reprioritize. No SSH, no database admin.

The metrics strip on the homepage of this site pulls from the same Supabase function that feeds the dashboard. You're looking at real numbers, not brochure copy.

What doesn't work (yet)

Cross-agent handoff. Getting the sourcer's output into the lister's queue without a human in the middle still has rough edges — schema drift between channels is the villain. The next project is a unified Product record that every agent respects.

Also: the content agents (YouTube, Fiverr responses) still need a human final pass. They're good enough to draft, not good enough to ship unsupervised. That's probably fine.

If you're building something similar

• One event log, everywhere. Every agent writes to the same table. Future-you will thank you.
• Health checks before features.You can't scale what you can't see broken.
• Self-healer early. The first version can be 20 lines. It pays for itself the first week.
• Boring orchestrator, clever agents. The cleverness lives at the edges, not the center.

More posts coming. If you're building a system like this and want to compare notes, get in touch.