Amazon Ops Agent

A busy household runs a surprising amount of commerce operations: orders arriving daily, return windows quietly expiring, refunds that need verifying against what was actually sent back. None of it is hard; it's just relentless, and missing a return window costs real money.

Off-the-shelf tools don't help because the state lives in two messy places: a mailbox full of order and delivery emails, and a retail portal built for humans clicking buttons. This is exactly the kind of friction agents should absorb.

The agent runs on my own hardware, deliberately: the mailbox is sensitive, so the data never leaves the house. A scheduled scan reads order, shipping, and delivery emails and turns them into structured facts: what arrived, when, and when its return window closes.

A deadline tracker walks those windows daily and escalates anything at risk. When a return is approved, browser automation (Playwright) drives the retail portal the same way a person would: initiating the return, selecting the drop-off method, and capturing confirmation. A final loop verifies the refund actually landed and matches the item. It also watches prices after purchase, and when automation hits a wall it escalates to Amazon's live support chat and negotiates directly with the human agent.

Guardrails are the design, not an afterthought: least-privilege credentials, human-approval checkpoints before anything irreversible, and every action logged so the system can be audited after the fact.

The agent has been running as part of my daily home operations: deliveries detected, windows tracked, returns filed with confirmations captured, refunds verified. Its best moment so far: it caught a price drop days after a purchase, opened Amazon's live support chat, negotiated with the human agent, and won the credit back automatically. The interesting result isn't the money saved; it's the proof that a bounded agent with narrow permissions can own a real, messy workflow end to end.

The dashboard on this page is a simulated replica of the real one, running on synthetic data, so you can watch how it works. For the war stories, ask The_D_Bot; it knows this project well.