I Gave an AI Agent the Family Calendar. Here's What Happened.

Five kids. Swim meets, marching band, picture days, permission slips, orthodontist appointments, and a weekly grocery run. For years the “system” was two inboxes, a shared calendar nobody trusted, and a lot of “wait, was that today?”

So I did what I do for my customers: I treated it like an operations problem and built an agent to run it.

What runs on its own now

School mail → family calendar. Every weekday morning an AI agent (Claude, running as a scheduled task) reads the last four days of school and sports email across both parents’ inboxes: district newsletters, Synergy blasts from teachers, band boosters, the swim coach’s weekly schedule. It pulls out only concrete, dated items — games, practices, concerts, no-school days, form deadlines — and writes them to our shared Google Calendar, tagged to the right kid. It doesn’t stop at the inbox, either. When a newsletter is just a link to a Smore page, the agent follows it and reads the actual newsletter. It scrapes the band program’s public web calendar directly, so rehearsal changes and performance dates land on our calendar the same day they’re posted — no waiting for someone to forward an email. This morning’s run added a middle-school open house, an elementary picture day, and a band-banner payment deadline. It also caught that the swim coach cancelled the default afternoon practice for Labor Day and removed it before anyone showed up to an empty pool.

Snap a photo, get a season. The paper calendar that comes home in the backpack — the band schedule, the swim season, the school-year “days off” grid — I take a picture of it on my phone and upload it. The agent reads it, extracts every dated item, checks each one against what’s already on the calendar, and adds the rest. A full season of events in the time it takes to photograph a sheet of paper.

A weekly recap. Once a week the agent sends a short summary of what it added, what it changed, what it skipped and why, and what’s coming up. It’s the digest that makes “set and forget” actually safe — I still know what the calendar knows.

A verified swim season, owned by the agent. Thirteen meets through the state finals plus a recurring practice block live on the calendar with provenance markers. If a coach’s email moves or cancels a meet, the agent updates the existing event rather than creating a duplicate — and tells me.

The Family Hub — on the fridge and in every pocket. The heart of it is a single-page web app on a home server. It runs full-screen on the Samsung Family Hub fridge in the kitchen and installs as a PWA on everyone’s phone, so the same screen is on the fridge door and in every pocket. One glance shows:

The agent refreshes the hub at the end of every run, so the fridge, the phones, and the parents are always looking at the same truth. It’s the piece the kids actually use — because it’s the one they can’t ignore on the way to the snack drawer.

Meals and groceries. The meal plan is a simple API on the same box; the agent keeps ~100 days of dinners populated and the fridge shows tonight’s. Then it turns the menu into a Walmart cart through a small Walmart cart service we built — an internal API that drives the ordering flow, handles Walmart’s bot-detection quirks, and gives the agent a clean “add these items” interface instead of a fragile screen-scrape. It defaults to store brand, but it remembers our preferences: Kraft mac & cheese, not the generic, because that’s the hill the kids will die on; the specific poppyseed dressing for a birthday salad. It claims every “Walmart Cash” offer it finds, re-verifies the cart because Walmart silently drops items — and then stops. It never checks out. A human reviews and pays.

The net effect: planning a week of menus and building the grocery order used to eat hours of a Sunday. It’s now a few minutes of review.

The rules that make it trustworthy

An automated calendar is only useful if the family believes it. So the agent operates under hard constraints:

Why this matters beyond my kitchen

Everything here is the same discipline I’d apply to a production system: idempotent runs, a ledger to prevent double-processing, explicit failure modes, human approval at the irreversible steps, and logs before dashboards. The domain happens to be permission slips instead of deployments.

The interesting part isn’t that an AI can read email. It’s that with the right guardrails, you can hand it ownership of a boring, high-stakes, never-finished job — and it stays quiet until something is actually worth your attention.

That’s the bar I hold for the automation I build for my customers, too.


Built with Claude (Anthropic) scheduled tasks and vision, Microsoft 365 and Google Calendar connectors, a self-hosted family hub, and a homegrown Walmart cart API. Happy to share the playbook — DM me.