Hey HN! Alex here. I built an AI agent that handles all the accounting at my company -- contractor invoices, payments in fiat and Bitcoin, bookkeeping, expense tracking, 1099s. It's called Sheila and it's open source (AGPL).

The short version: I tried building complex multi-step automation in code first. It was fragile. What actually worked was building 50+ granular scripts that each do one thing (check a balance, send a payment, append a row to a spreadsheet, read an email), then writing a 600-line AGENTS.md file that describes the workflows in plain English. The AI agent reads the instructions and chains the scripts together.

The key insight for me was that the human feedback loop is what makes agents actually work. I sit in the terminal watching Sheila process invoices. She drafts emails and prepares payments, I see what she's doing before it goes out. Hundreds of iterations between me and the AI is what made this reliable -- not some autonomous agent platform where you click "deploy" and hope for the best.

I think the industry has this backwards right now. Everyone is building agent platforms top-down. But you can't design a useful agent abstraction layer without first understanding what a useful agent looks like, and you can't understand that without building one from scratch, script by script.

Source: https://gitlab.com/soapbox-pub/sheila

Happy to answer questions about the architecture, the Bitcoin payment flow, or why I think most agent frameworks are solving the wrong problem.