How Norma Gets Things Done: Inside the Harness of a Personal AI Assistant

Think of Norma as a very talented new assistant on their first week. Brilliant, fast, occasionally a little too eager. You wouldn't hand them your credit card and walk away. You'd give them:

• a clear list of what they're allowed to do,
• a rule to ask before anything that can't be undone,
• a notebook where they write down everything they did,
• one door to the outside world, with someone checking who goes through it.

That's the harness. Now the slightly longer version.

Every message you send goes to one capable model that can use tools: set a reminder, look something up in memory, list what's on your plate, cancel, move, undo. It can use a handful of tools per message, within a strict time and cost budget, so a simple “remind me to call mum at 6” never turns into a 12-step philosophical journey.

The important rule: the model proposes, the code decides. The model says “I'd like to create a reminder for 18:00”. Norma's own code checks it, does it, and hands back the real result. Norma's reply is written from that result, not from the model's imagination. So Norma can't tell you something is done when it isn't. We check that on every reply.

Some actions are easy to reverse. Some are not. Deleting memories, cancelling everything, sending an email or changing your calendar all get the same treatment: Norma tells you in plain words what it's about to do, and waits for your yes.

And if you change your mind afterwards? “Undo that” works, and “why did you do that?” gets a real answer, straight from the record of what happened.

Behind the scenes, every step Norma takes is written to an activity log: your message, what the model was asked, which tool it used, what came back, what you confirmed. Nothing is overwritten, only added.

That sounds boring. It's the opposite. The log is what lets Norma:

• explain itself (“I moved it because you said Thursday”),
• pick up exactly where it left off if a server restarts mid-thought,
• wait for things that take days, like a reply from a colleague,
• and let us find and fix mistakes instead of guessing.

And yes, it respects forgetting. Ask Norma to forget something and it disappears from the log too, not just from memory.

When Norma starts touching things outside itself (your calendar first, email after that), every single action goes through one gatekeeper we call the Capability Gateway. Before anything leaves, it checks:

• Is Norma allowed to do this at all?
• Did you say yes, if it needs a yes?
• Did the instruction come from you, or from some text inside an email or web page? (Those are treated as information, never as orders. Sorry, sneaky newsletter.)
• Has this already happened? If a connection drops right after a calendar event was created, Norma checks instead of blindly trying again. No double-booked dentist appointments.

Your passwords and access keys live behind that door too. The AI model never sees them.

Real people don't text in paragraphs. They text like this:

hey
can you
move my 3pm to tomorrow

Norma now waits a moment and treats a burst of messages as one thought, so you get one good answer instead of three confused ones. It even waits a little longer while you're still typing. Small thing, huge difference in how it feels.

Not every step needs the most expensive brain. Deciding “is this a question or a request?” can be done by a tiny, fast classifier. Looking something up needs no AI at all, just a database. The big model is saved for the moments that need judgment. Every change we ship has to keep quality at least as good while keeping cost at or below where it was. That's how we plan to make a capable assistant affordable for everyone.

Built and running in our builds:

• One model with tools, where the code decides and replies come from real results
• Ask-first for anything you can't undo, plus “undo” and “why did you do that?”
• The activity log of everything Norma does, forgetting included
• The Capability Gateway, tested against a practice calendar: no secrets reach the model, no duplicate events after a dropped connection, nothing left behind when you disconnect
• One reply to a burst of messages
• A cost budget on every run

Building now:

• Measuring how good the harness really is, not how good it feels: response time targets (most replies within about 3 seconds), and doing the same task right several times in a row, not just once
• Clear examples on every tool so Norma picks the right one, in Swedish and English

Next:

• Tools that remember. A restaurant search that already knows your allergy. And Norma remembering what it did, not just what you said.
• Tasks that wait. “Keep an eye out for Sarah's reply and nudge me Friday” survives restarts and wakes up on its own.
• Your calendar, then your email, all through the gateway, nothing sent without your say-so.