Daniel Pyrathon (@pirosb3) on X

X (formerly Twitter) ·

2 min read Original article ↗

One of the most sticky agent workflows I’ve built orders groceries for our family.

While I love picking my produce at the store, grocery shopping in my household is annoying in a very specific way. It requires coordination and often leads to suboptimal decisions (I’m a sucker for chocolate and chips section).

My wife and I remember things at random moments. Sometimes it’s obvious: bananas, yogurt. Sometimes it’s vague: ingredients for carbonara, something easy for dinner, or salami and a cheese that pairs well under $10. Sometimes we get pulled into side quests: “ooh that pasta is on discount”.

So I built an agent for it.

Every other day at 9am, Lou opens a new cart for @unionmarket, checks what’s already there, and emails me and my wife that the cart closes at 2pm. We reply with whatever we want. The interface is just email via @agentmail.

When one of us replies, Lou uses @browserbase to open the grocery site, search, choose reasonable items, and add them to the cart. “Ingredients for carbonara” becomes the likely set of items. “A cheese that pairs well with salami, under $10” becomes a sensible pick. If brand or size matters, Lou escalates instead of guessing. I also told it to bias toward lower saturated fat options.

At 2pm, Lou checks the total:

Over $100: escalate to me

Between $10 and $100: submit the order, since the card is already stored

Under $10: wait for next time unless we said it’s urgent

This has been a real quality-of-life improvement. Groceries are mundane, recurring, and just annoying enough to create friction every week. Lou gives us that time back. I've also saved money on unnecessary sugary purchases.

That’s why I think this is a good example of what agents are actually good for.

The value is end-to-end ownership: wake up on a schedule, coordinate with humans, operate a browser, remember what happened earlier, make bounded decisions, and escalate when needed.

This is the idea behind Memento and the agent systems I’ve been building: useful agents aren’t one-shot prompts, they’re durable workflows. They need to sleep, wake up, receive new information, continue later, and preserve intent across time.

My grocery agent isn’t AGI. It just removes a chore I dislike several times a week. That’s a pretty good value prop!