The AI agent that buys my groceries and runs my apartment

At five in the morning the cat wakes me up — breakfast time. Then the litter box, the blinds, the lights, the air conditioning. Five remote controls used to sit on my desk, and one day that finally broke me.

Fifteen minutes every morning adds up to 91 hours a year.

Once a week there’s the shop on top of that: get ready, drive, sit in traffic, pick everything out, pay, carry the bags up. Two hours. Another 96 hours a year.

Together that’s almost a month spent on things nobody has to do.

Today one remote is left in the flat — the Apple TV. Lights, blinds and climate run themselves; I haven’t touched a switch in a year and a half. An AI agent orders the groceries, leaving me five minutes to check the basket and a walk down to the door. The cat still wakes me at five. Automation lost that one.

If an AI agent still sounds like magic from the future, here’s what it’s actually made of. Five parts, no magic in any of them.

Part	What it does
1. Model	Thinks and answers
2. Persona	Who the agent is and how it behaves
3. Memory	What it knows about you
4. Tools	What it can actually do
5. Self-improvement	Why it keeps getting better

1. The model — the part that thinks

A language model was trained to do one useful thing: guess the next word with high probability. It guesses a word, then the next one — and an answer assembles itself. It doesn’t love you, doesn’t understand you and feels nothing. Underneath there is math and a table of weights.

That doesn’t make it any less useful. Once you understand the mechanism, you stop waiting for miracles and start operating a tool.

A token is the chunk of text a model works with — roughly four characters in English. Both volume and price are counted in tokens.

2. The persona — who the agent is

Picture an adult with an empty head. He can speak and reason, but knows nothing: not who he is, not what he does, not a thing about you. You, as his maker, put in what belongs there — with one condition: don’t overload him. That condition is financial, and I’ll come back to it.

You start with a sense of self: who he is, what he does, what tone he speaks in, what’s off limits, how to address him, whether he greets people when joining a call.

[SCREENSHOT: persona settings — aliases, role, tone of voice, constraints]

The rules and constraints field matters most in practice: without it the agent agrees with everything you say.

Voice is configured in the same place. Agents stopped being text-only long ago, and talking turned out to be faster than typing: a Stanford study measured a threefold gap against phone typing — 161 words per minute with 20% fewer errors.

[SCREENSHOT: voice picker, tone and speech rate]

3. Memory — what the agent knows about you

The most complex of the five parts, and nobody has finished solving it. It can be assembled in wildly different ways. Here’s the baseline worth starting from:

Type	What it holds	Example
Facts	What the agent learned about you	«Henry doesn’t like fish», «Henry likes South African wine»
Lists	Any collection	Groceries, clothing, smart home devices, rules
Goals	Where you’re heading and by when	«Learn English by December», with milestones

The difference shows the moment enough facts pile up. Ask an agent with no memory to order wine:

Found 1,532 bottles. Which one would you like?

The same request to an agent that knows you:

Your South African is in stock at the shop nearby. Tell me when you need it and I’ll arrange delivery.

Say «tomorrow, for a meeting» and it checks the neighbouring shops in case the same wine is cheaper, then books a courier ahead of time on the economy option. Say «urgently» and it takes the priority slot.

[SCREENSHOT: memory tab — facts about the owner]

[SCREENSHOT: memory tab — goals with milestones and deadlines]

This is also where «don’t overload him» comes in. You pay per token: at 30,000 tokens a request costs pennies, at half a million every message becomes a heavyweight — and you pay full price even when you asked about the weather. So memory lives outside the conversation, and only what’s needed gets pulled in.

4. Tools — how the agent acts

A model with memory can hold a conversation. To do things, it needs hands.

Tool	What it is	Example
Skill	An instruction for how to do something — a text file with steps	Compare prices across three shops, build the basket, show it to me
Connector, MCP	Access to a service through an open standard	Shop, messenger, smart home, calendar

Here’s what that looks like in practice. I buy from three shops: vegetables from one, meat from another where the quality is better. The agent compares prices and splits the order. The butcher only takes orders through Line, so the agent messages them and agrees on timing. I pay, it books a Grab courier, the order gets collected and delivered. If I’m out, it’s left at the front desk.

One benefit surfaced by accident. Slow delivery costs a fraction of express, but people don’t book it — nobody wants to wait ninety minutes. The agent doesn’t mind. It plans ahead and takes the cheap slot.

5. Self-improvement — why the agent grows

Every few hours the agent goes through accumulated conversations and pulls out new facts, goals and rules. All of it lands in memory, and the next conversation runs with it. It watches its own spending the same way: every day it reviews where the money went and tightens its own limits.

That’s the difference from an ordinary chat window. ChatGPT is built to serve everyone at once, so its memory is deliberately shallow. Your agent is configured for you alone — and the agent you had a month ago is noticeably worse than today’s, running the exact same model.

What it costs

What	Price
Household agent	$5–10 a day, around $150 a month
Coding agents	$100–200 a month on subscription

The top subscription gets you a couple of developers working almost around the clock. Compare that to one salary and it becomes clear why I stopped hiring.

People ask about mistakes a lot. The most expensive one I’ve seen wasn’t made by a machine: an employee wrote a function that called itself, crashed and restarted, hitting a paid Google API every cycle. Six weeks and $200 before anyone noticed. An agent can overspend too — the difference is you can cap it, and it polices the cap itself.

What’s next

Five parts is the map. Next I’ll go through each one:

• how search by meaning works: vectors and embeddings in plain language
• what a request costs and how to count tokens
• how to write skills and wire up MCP
• how to give your agent a voice

If you want an agent like this and don’t want to spend months figuring it out, I run one-on-one training. Get in touch.