2026-03-08 · 6 min read

The Agents That Run My Mornings

I'm a product manager. I have never shipped a line of production code in my life. And yet by the time I wake up most mornings, an AI agent has already written my daily briefing, pulled the weather, scanned my calendar for conflicts, and flagged the three emails that actually need me.

I didn't hire anyone. I wired this up myself, over a few weekends, by describing what I wanted in plain English and letting the model write its own plumbing.

I want to tell you what that's actually like, because most writing about AI agents is either breathless or terrified, and the reality is neither. It's quieter than that, and stranger, and a lot more useful than the demos make it look.

The setup

The setup is boring on purpose. A handful of small agents, each doing one job, each on a schedule. Seven run right now. A few more are half-built.

At 6:30 one assembles the morning brief: the day's calendar, the weather, what I said I'd do, what's overdue, the two or three things that actually matter before noon. At 7:00 another empties my inboxes, sorts what came in overnight, and drafts replies to the ones that need a reply. At 7:05 a small one digs up what I was doing on this date in past years and drops it into today's note, which is a nice way to be ambushed by your own history over coffee.

Then the rest of the week fills in around the edges. Monday morning one scans the week ahead, every meeting and commitment for the next seven days, and flags the conflicts and double-books before I walk into them. Wednesday it refreshes a dashboard of everything I'm trying to publish and how long each piece has been sitting. Friday at 9pm one writes me a letter about the week, in my own voice, because before it writes a word it reads a couple of my old essays to get the texture right. And at 11pm every night the quiet one runs: it walks through everything I wrote that day, updates a map of the projects and people I'm tracking, and checks that all the others are still healthy.

They run on my own laptop, on a timer, whether or not I'm paying attention. No backend. No team. Nobody with an engineering degree, least of all me. Just instructions, written carefully once.

How a non-engineer wires this up

This sounds harder than it actually is.

macOS has had a scheduler built into it for twenty years. It's the thing that already wakes your computer up to run backups and check for updates. I just gave it new jobs. Each agent is a plain text file of instructions and a one-line script that, at the scheduled time, hands those instructions to Claude on the command line and lets it work. That's the whole trick. There's no server I rent and no cloud bill, and there was no tool whose internals I had to learn. The scheduler was always there. I pointed it at a model.

I didn't learn to code to build this. I learned to write very clear instructions, and the model wrote the code. And the more I do it, the more I think the skill that actually transferred wasn't engineering at all. It was knowing exactly what I wanted, in enough detail that a literal-minded machine couldn't misread me. Which, when I say it out loud, is just the job. It's the same thing I've been doing for engineers my whole career. The only difference is that now there's no handoff, no sprint, no waiting two weeks to find out I described it wrong.

The one rule that makes it safe

There's a single rule that turns this from a fun toy into something I trust with my actual life: nothing sends.

Every agent drafts. None of them act on the world. The inbox agent writes the reply and leaves it sitting in my drafts with a label that means "a human still has to look at this." The agent that scans my week proposes a fix for an overbooked Monday; it doesn't touch the calendar itself. An agent that can do things while I'm asleep is a completely different risk category from one that just prepares things for me to approve when I wake up, and I decided early that I only wanted the second kind. The supervision isn't a limitation I'm working around. It's the design.

Because the agents are wrong sometimes, and they are wrong with total confidence: the most dangerous artifact the system produces isn't a missed email, it's a polished, plausible draft reply that has the facts subtly backwards and is one click from being sent. The nothing-sends rule exists precisely so that draft has to pass my eyes first.

The other half of trusting it is being able to see what it did. Every agent leaves a footprint: a line in today's note saying it ran, a row in a little health dashboard, a log I can read if something looks off. If an agent fails, it says so and keeps going instead of silently dying. So at any moment I can glance at one page and see which agents fired this morning, which ones choked, and what each of them touched. The system watches itself, and it tells on itself. That, it turns out, is the difference between automation you trust and automation you quietly stop relying on.

The uncomfortable part

Here's the part nobody warns you about.

Once you have agents watching your life, they get very good at noticing what you are not doing.

My publishing dashboard knows, to the word, that I drafted zero words last week, and exactly how many days each unfinished piece has been sitting untouched. The agent that walks my notes at 11pm keeps a quiet running map of the projects I've stalled on, and that list is never as short as I'd like. The system didn't make me more productive. It made me more accurately aware of my own avoidance, which turns out to be a different thing, and a more uncomfortable one.

The leverage was never that the agents do the work. It's that they quietly remove every excuse except the real one. The chaos used to be the cover. When my inboxes were a mess and my projects lived in my head, I could always tell myself I was busy. There's nowhere to hide from a system that takes attendance. It doesn't nag, exactly. It just keeps an accurate record, and an accurate record turns out to be its own kind of pressure.

If you want to try this

So if you want to try this and you can't code, don't start with the dashboards. Start with one agent that does one thing you'd otherwise skip, and let the model write the script while you watch what it surfaces.

The starter ritual:

Pick one job you'd otherwise skip: a fifteen-minute morning brief, say.
Write it in a paragraph, in the same plain words you'd use to ask a sharp assistant. Let the model write the script.
Make it draft, not send.
Make it leave a note when it runs.
Watch what it surfaces for a week before you build the second one.

The first genuinely useful thing it tells you will probably be something about you.

That's it. That's the post. For weeks the dashboard has reminded me I wasn't writing. The agent did its part this morning. This time, so did I.