Sheila the accounting agent pays contractors, but only after a human OK
Invoices, contractor payments in cash or Bitcoin, bookkeeping, 1099s. Sheila drafts and prepares it all; the founder watches and approves before anything goes out. Hundreds of iterations made it reliable.
Done withCustom agent✓ Verified· by HN user alexgleason· Oct 15, 2026
Result Reliable after hundreds of iterations
“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”
The project's public code: a contractor-payments agent built and tested with Claude Opus 4.6, whose instructions make it ask the person to confirm before it acts on payments and expense claims.
Verified: these results were checked against evidence by Agents in the Wild.
What they did
Built Sheila: agent handles contractor invoices, payments (fiat/Bitcoin), bookkeeping, 1099s; human watches terminal and approves before actions execute.
How it went
Hundreds of iterations between human and agent made it reliable; key insight is human feedback loop, not autonomous operation.
Try it yourself with Custom agent
Build an accounting agent that: reads contractor emails, extracts invoice data, processes payments, maintains records. Implement human approval loop: agent prepares actions, waits for human confirmation before executing. Focus on reliability through iteration.
A starting prompt written by Agents in the Wild. Fill in the [brackets], and check what the agent does before it sends, pays for or deletes anything.
Vet bills mean a paper chase of card charges, emailed invoices and insurer forms. A homemade agent matches each charge to its Gmail invoice, fills the claim form with Gemini's help and stages a draft. A person hits send.
A Raspberry Pi rides in the car and runs a local AI model that answers spoken questions through the car's own speakers, even with the internet switched off. It also posts engine faults into the household chat room.
Gmail, calendar, maps, contacts, alarms: their homemade agent controls them all, and even taps through Telegram for them using the phone's accessibility service. All data stays on the phone.
Too much news, too little time. They built an agent that sends a daily summary so nothing important slips past, and that digest is what pointed them to the thread where they shared it.
A home service business connected Grok Bot to its phone system after a few hours of setup. It now listens to every call and posts a summary to the team's Slack channel.
Every Monday a 12-tab marketing tracker got a 30-minute scroll and a vague gut check. Claude Cowork read every tab and built a color-coded dashboard inside the live Google Sheet.