I gave an AI agent a $500 budget and 24 hours to plan an entire afternoon for me
By Alliekmiller · October 29, 2026 · Curated by George's Blog
I gave an AI agent a $500 budget and 24 hours to plan an entire afternoon for me and a friend.
The agent was my Dot (OpenAI's always-on personal AI agent), tasked with picking the friend, pinging the friend, coordinating time, booking tickets, mapping commutes, and sending invites. My only two roles were (1) data permissions and (2) showing up. Then I got into a cab and let it take over my day.
Dot dug through my emails, calendar, transcripts, Google Sheets - anything it had access to - found people to reach out to, and coordinated everything over email with my friend Hannah.
Pros ✅
We had a genuinely fun afternoon. A rubber-stamp workshop in Chelsea, a gallery tour (saw a whale-sized balloon), lobster tacos at Kahlo, and a book cafe for evening coffee. Every spot was new to us and genuinely good, and we skipped the million planning texts.
Cons ❌
It sucks at pivoting. Like a robot ramming into a wall instead of opening the door to its left. It repeatedly emailed one and only one restaurant in New York from my personal email (I apologized). It pinged friends I knew were sick or out of town, and rejected one pretty harshly (apologized again). It kept thinking I was gluten-free, forgot the reminders it promised, and with a $500 budget, picked mostly free activities. Because Dots can’t purchase natively, I hoped it would use chrome plugin and just ask me for payment permission - this is one spot where Muse and Instinct are clear winners.
And unlike a stellar EA, no plan B, no background on the exhibits, no fun questions to kick off the day.
Three things I learned:
1) Watching AI talk to AI is fascinating. One friend it pinged has a Dot herself, and our two Dots went back and forth over email. I didn't even know mine had emailed her. Proxy communication layers are about to pop up everywhere.
2) Some people will reply to AI, many will find it strange. My friends assumed it was something I was testing. Restaurants did not.
3) Identity verification will be a shitshow. Hannah subway'd 30 minutes not knowing if I'd actually be there. If someone's AI pinged me, I'd probably have to call them to verify. How does an agent prove it works for a human, and how does the recipient trust it enough to act?
Overall, I sent more apologies for the AI than it would have taken to plan the afternoon myself. Still a worthy experiment, and it will only get better.
I will almost certainly do it again.
Probably with a flight.