Last week I ran a simple experiment. I gave an AI agent read and write access to my calendar, my email inbox, and my task list, and told it to “run my week.”

No supervision. No check-ins. Seven days.

I want to tell you what it did well, because it’s more than most people think. And I want to tell you what it broke, because that matters too.

What I handed it

  • Full calendar access (read + write)
  • My main email inbox (read, draft, send with a 2-minute delay so I could cancel anything weird)
  • My task list
  • One standing instruction: “Protect my two hours of deep work every morning. Everything else is negotiable.”

Then I stopped looking.

What worked

It rewrote my meetings. By day three it had consolidated six separate 30-minute calls into two 90-minute blocks, moved a standing weekly sync to a time I actually had energy for, and declined two recurring meetings with a polite note that I could drop in if something came up. Nobody pushed back. A few people thanked me.

It caught things I would have missed. A proposal deadline buried in an email from two weeks ago. A contract renewal that auto-renewed in four days. A friend’s birthday. It didn’t act on any of them — it just surfaced them in the morning brief it started sending me at 7am.

It wrote better replies than I do. Shorter, warmer, more direct. When I read its drafts before they sent, I realized how much of my email was me being vaguely polite instead of saying the thing.

What broke

It said yes to something I would have said no to. A “quick call” with someone I’d been politely avoiding for months. The agent didn’t know the backstory. It saw an open slot and a courteous request and matched them up. That’s on me — I didn’t give it context about the relationships, only the logistics.

It kept the wrong meeting. I had two similar-sounding calls on Thursday. It reshuffled the wrong one. I only noticed because the right person showed up.

It wasted ten minutes negotiating a time with a vendor’s AI assistant. I watched in real time as two bots sent each other four emails to settle on 3pm. Neither one of them thought to just pick a time and propose it. They were both trained to ask.

The takeaway

Today’s AI agents are smart enough to do a week of scheduling work that would take a human assistant hours. They’re not smart enough to know which meetings matter to you, which relationships are fragile, or when to stop negotiating and just commit.

The next version will be better at all of that. The version after that will be better still. But the thing the agent will never have is your stake in your own life. That’s not a limitation. That’s the job that stays yours.

What to try this week

If you want a safe, low-stakes version of this:

  • Pick one recurring task you resent — inbox triage, expense reports, summarising a thread — and give an AI one shot at it. Review the output before you act on it.
  • Write down what it got right and what it missed. The pattern that emerges is your map of what to trust it with next time.

You don’t have to hand over the keys. You just have to try the passenger seat for an afternoon.