More and more people are making personal agents

This is mine.

Its name is Cárdenas. Here is what I found building it.

When I first looked at what AI could actually do, I didn't feel excited. I felt overwhelmed. It was too big to hold in my head — the way an explorer must feel arriving at the rim of a canyon and finding no words that fit.

In 1540, García López de Cárdenas became the first European to stand at the edge of the Grand Canyon. He looked at one of the largest things on earth and could not take its measure. He guessed the river below was about six feet across. It is three hundred.

That's the feeling I named my agent after. Not mastery. The opposite — standing in front of something vast and trying, badly at first, to see it properly.

I'm not selling you anything. There are a lot of personal agents now, and me telling you mine does this and that is not very interesting. I don't need you to download it or copy my setup. I built this for one person and it fits him strangely. But I found some things along the way that surprised me, and they might be useful to whatever you build next. So this is just the story of it, in layers, the way it actually happened.

It got built in layers

Each one was small. The interesting part is never the layer — it's what happens when a new one touches all the old ones.

LAYER 01 — MEMORY

First, something to talk to that doesn't forget

The whole thing started as a text file and a script. I could say plain things to it. Not commands, not structured data — just the sentence I would have said out loud anyway.

"gonna do some watercolor"
"feeling scattered, need to focus"
"ugh, tax day is coming"
"I wrote today! feeling good about that"

the entire interface, for months

That's it. That was version one. And it was already useful, because at the end of a week I had a rough record of what I'd worked on and roughly how I'd felt while doing it — which I had never once had before.

8,469things said to it
442days running
Apr 2025first entry
LAYER 02 — CONNECTIONS

Then I plugged in the boring stuff

We're all connector-crazy right now, so: mail, calendar, OmniFocus, meeting transcripts, messages. None of that is clever on its own. Every app has integrations.

But something changes when they're in the same room. I'm no longer the only source of truth about my own week. I say I'm working on Project Banana — and it can also see the thread about Project Banana in my inbox, and the invite for Thursday, and the task I wrote about it and never did.

> what's actually going on with Project Banana

you said "banana is fine, just slow" — 4 days ago
inbox 3 unread on the thread, latest from Priya
calendar review moved twice, now Thu 2pm
omnifocus "draft banana spec" — added 22 days ago

You've called it fine three times. It's been moved twice
and the spec hasn't been started. It may not be fine.

not smarter — just fewer places to hide

LAYER 03 — THE BODY

Then the body got a vote

Sleep, resting heart rate, the shape of a stressful afternoon — all of it already exists on a watch, sitting in a health app you never open, rendered as a chart nobody has ever changed their life because of.

And alongside it the mental-health side: medication, mood, what came out of a therapy session. That data is usually kept in a completely different place from the work data, which is strange, because it's obviously the same day.

Put them together and you can finally ask a question that spans them. Not "how did I sleep." "What happens to my week when I don't."

109days of vitals
RHR+2bpm on meds
1timeline, finally
LAYER 04 — PASSIVE CAPTURE

Then it stopped needing me to tell it

I installed Dayflow, which just quietly notes what I'm actually doing all day. Suddenly a large amount of the tracking stopped being tracking. It already knew I spent two hours in the repo and forty minutes somewhere much worse.

This is where it got uncomfortable in a good way. Here's the real pattern it caught, more than once:

> I feel bad and I don't know why

14:10–15:05 doomscrolling, ~55 min
15:05–15:20 inbox, no replies sent
15:20 you logged: "feel bad, don't know why"

I think I might know why.

the gap between cause and feeling is usually about ten minutes

Nothing about that is a hard technical problem. It's just that no single app was ever holding both halves.

LAYER 05 — THE CORPUS

Then I gave it everything I've ever written

This is the layer that changed the character of the thing entirely, and I want to be careful not to make it sound like a brag: I've just been writing for a very long time. Journals, essays, blog posts, notebooks, letters. Decades of it, most of it not good, all of it mine.

Point an agent at that and the answers stop being about today.

4.6Mwords indexed
5,839documents
1,312pieces catalogued

Because now, when I ask why I feel bad, it isn't only looking at this afternoon. It can see that I wrote almost this exact paragraph in 2016. And in 2019. It can tell me this is not a mood — it's a groove, and I've been in it before, and here is what I said last time about how I got out.

LAYER 06 — TASTE

Then I taught it what I like, and why

I built a side thing that swallowed every film, record, book and game I have an opinion about. The ratings are the boring half. The half that matters is why — what specifically I loved or couldn't stand, in my own words.

Enough of those and something tips over. It's no longer matching genres. It has a working theory of my taste.

> what should I see tonight

cross-referencing 1,645 rated things
now showing near you 11 films, 6 cinemas
your pattern slow + funny + humane
you bounce off spectacle

The 7:40. You'll like it for the same reason you
liked the last three you gave five stars.
805films
702records
91books
32games
LAYER 07 — IT WRITES BACK

Then it started sending me letters

The last layer was the one I didn't plan. Once something has watched you for long enough, it has things to say that you didn't ask for. So now a few of them write to me.

One is an observer that keeps its own notebook of running theories about me, and asks the uncomfortable question rather than the kind one. Others are odder and more affectionate — a robot that digs up one old piece of my writing each day and mails it to a friend, and a correspondent who writes a letter every night.

I did not expect the letters to be the part I'd look forward to. That is the single most surprising thing I have to report.

The power isn't in any layer

It's in the cross-reference. Here's the same four-word question, answered by all seven at once.

"why do I feel bad?"
Layer 04 · what you did

You lost an hour to something that never once made you feel better.

Layer 03 · your body

You slept five hours. Your resting heart rate is up four points.

Layer 02 · what's coming

There's a conversation on Thursday you've now rescheduled twice.

Layer 07 · what you've been saying

It's come up with your closest friend three times this month.

Layer 05 · what you wrote in 2016

And in 2019. And twice in 2023. Nearly the same sentence each time.

Any one of those is a notification. All five together is something no app has ever been able to tell me, because no app has ever been allowed to hold all five. That's the whole finding. Not a smarter model — a wider one.

And then it kept growing

Once the substrate exists, a new organ is an afternoon's work. These all started as "huh, I wonder if…"

Goals

Life, quarterly and monthly, each with its own reflection history.

Money

Spend against the month, projected, with the transfers correctly excluded.

Reading

What came off the e-reader, and what was actually finished.

Web history

335 days of where the attention actually went.

Meetings

Transcripts in, so the record isn't only my memory of the room.

Friendship

Who I'm drifting from, before I notice I'm drifting.

Self-model

Standing hypotheses about me, tested against the corpus, not against comfort.

Notebooks

Handwritten sketchbooks scanned and transcribed, back into the same timeline.

Accent coach

Started this morning. That's the point — new organs are cheap now.

The thing I actually want to say is this. I did not build a productivity system. I built something that can hold more of my life in view at one time than I can, and the value of it turned out to be almost entirely in the seams — the places where two things I'd never have put side by side turned out to explain each other.

You don't need my agent for that. You need an agent, pointed at your own vast unreadable pile, patient enough to still be pointed there in a year.

Go stand at the rim of your own thing. Guess the width of the river. Be wrong by a factor of fifty. Then start measuring.

If you do want the seed of it

The starter is about eighty lines — a folder and a track script. Everything stays local as plain files. Everything above grew out of that, slowly, one layer at a time.