Essay · August 14, 2026
Running production agents on a $30 server
A lot of what the industry bills as cloud complexity is markup on a problem most small firms don't have. The automation stack we run for a client, the agents, the workflow engine, the model gateway, the database, the observability, fits comfortably on a single server that costs about thirty dollars a month. Not a toy. The real thing, in production, isolated per client. Understanding why that's possible, and where it stops being enough, is worth more than any provider's pricing page.
Where the 10x bill comes from
When a competitor quotes cloud costs at ten times ours, they're usually not lying. They're passing through the bill from managed services priced for enterprises that need them: autoscaling nobody's traffic requires, redundancy across regions a regional firm will never touch, per-request pricing on models that a hard budget cap would have solved for a few dollars. For a firm of eight running a handful of workflows, most of that is insurance against a risk they don't carry.
What actually fits on one box
A modest server runs it all: the workflow engine, a model gateway that enforces a hard spending cap per client, a Postgres database doing real work, and the logging that tells you what every agent did and what it cost. The model spend is the only cost that can quietly grow, and a cap plus per-client logging turns that from a surprise into a number you watch. Each client gets their own isolated box, so no client's data ever shares a machine, a database, or a key with another. Isolation you can explain in one sentence is worth more than a compliance diagram nobody reads.
The honest limit
Here's where a single cheap server stops being the right answer, and we'll tell you before you're on the wrong side of it. One box is one failure domain: if it goes down at 2am, the system is down until someone brings it back, and the backups are only as good as the last time we tested a restore. That's an acceptable trade for a Monday briefing and an invoice router. It is not acceptable for something that has to answer every second with money on the line. When a workload genuinely needs high availability, it needs more than one box and a bigger bill, and pretending a thirty-dollar server can carry it would be the expensive kind of cheap.
We start small on purpose, because most work never needs more, and the work that does will tell you clearly when it's time.