Agents, tool calls and the plumbing that makes a model genuinely useful inside a real system.
Every week the same few volunteers spent two hours or more chasing sermon slides, copying Bible verses and fixing line breaks. So I built them an agent that does it on Thursday evening, for about seven cents, and still lets a person say no.
How the Sunday slides agent proves it never changed a verse, why Claude only gets to propose design changes, and how the design pass went from 50 cents to 8 cents a build.
Two Railway services, one cron, a webhook, a watchdog and a cost log. The real numbers from the Sunday slides agent's first weeks in production, including the Thursday it failed.
I assumed the model would use the tools sensibly. It did not. What changed once I started designing tools the way I design an API: narrower inputs, typed arguments, and error messages written for the model rather than for me.
Less about the hype, more about a specific kind of repetitive work I would like a program to take off my hands.
A month of letting an assistant write most of the first draft: where it saved hours, and where it cost me a weekend.
Next
I am always up for a conversation about architecture, trade-offs, or a bug that refuses to die.
Start a conversation