I handed my 2001 batch reunion to an AI agent, and I've mostly become the mailman
I got hold of Instinct about two months ago. For the first few weeks it was an add-on. Something I opened when I needed a taxi booked, a flight price checked, or a bit of personal finance tidied up. Useful, but peripheral.
Somewhere in the last month that changed. It's now the agent I reach for first, for actual personal work. And the thing that pushed it over the line wasn't a clever feature. It was a college reunion.
We're planning the reunion for the batch I passed out with in 2001. Twenty-five years on, most of us haven't been in the same room since. Somewhere in the middle of organising it, I stopped treating Instinct as a tool and started treating it as a man Friday. What surprised me is that almost none of the use cases were designed. We found them by pushing on the edges of what it could do, and it kept not falling over.
Here's what it ended up doing, in case you're planning something similar.
1. Registration, payment tracking, and a payment gateway that doesn't exist
I built a registration page through Google AI Studio. I instructed Instinct, and it instructed the Google agent to do the work. That chain took me a while to get comfortable with, but it works.
The interesting part is the payments. We don't have a real gateway. It keeps track of the payment details attendees submit and helps reconcile registrations against the bank statements I provide. It's a pseudo gateway, held together by validation rules rather than by Stripe.
It also posts a daily WhatsApp summary of new registrations and nudges the fence sitters. That summary comes to me, and I forward it on. Which is the honest description of my role in this operation.
2. Engagement, and a leaderboard nobody has to ask me about
A reunion doesn't work if the first contact people have with each other is the day they arrive. There needs to be something happening beforehand.
So Instinct built a leaderboard for the quizzes and games running in the group. I gave it the goal and the infrastructure once. It doesn't ask me about it daily. It knows the task, refreshes the page, and sends me a notification when there's something worth amplifying.
That last part matters more than it sounds. The value isn't that it does the work. It's that I don't have to hold the work in my head.
3. Surveys and matchups
It creates ad hoc surveys when we need to understand what the alumni actually want, and analyses the responses in real time.
The one I like most is travel. It matches people who'll land at the same airport around the same time, so they can share the transfer to the institute. Nobody instructed it to do that. It understood the goal and worked out that this was a way to serve it. It confirmed with me before acting, which is the behaviour I want.
4. Email campaigns
A reunion involves a lot of email. Students, professors, the institute, people who haven't replied to anything since 2003.
I gave Instinct the objective and suggested some infrastructure it could follow. It set up its own email and data storage to execute the plan. Every email goes through my approval first, without exception, occasionally to my mild dismay. But that approval step is the reason I'm comfortable letting it near a list of people I care about.
5. Artefacts
Infographics, announcement posts, multimedia for the group. Songs and videos too, to keep people engaged between the announcements.
I'll be honest that this is the part where I check the output most carefully. Generated artefacts are the easiest place for something to look fine and be wrong.
Two things I've learned
I have no idea which LLM Instinct runs on. I've never needed to know. What I know is that I enjoy working with it, and that the experience of working with it is what makes it good. Which is where I think agents are heading: the model becomes invisible, and the power moves into the UI.
The safety advice is not optional. Share carefully. Don't hand over your personal accounts; set up separate accounts for the agent to operate in. Give it the data it needs to know and nothing else. Financial data is a hard no for me.
Then add guardrails. Limit what it can reach and what it can do while it's in there. I spend real time thinking about where it could deviate, and when I'm in doubt, I ask the AI itself how it could get lost. That sounds odd until you try it. It's usually quite good at telling you the ways it could go wrong.
None of this required a clever prompt. It required a clear goal, a bit of infrastructure, and a willingness to keep pushing until something broke. Nothing has broken yet.
AI Learning Pilot
Bring agentic AI to your classroom
TechSambad's AI Learning Pilot puts agentic AI in front of your students for one module, not a semester: a live 90-minute taught session inside your timetable, an AI-built pre-class reading brief mapped to your module, a question bank with grading rubrics, a student takeaway pack, and a written outcome report. Run it once and judge it on what your students can do afterwards — Rs 2,000 + GST for the college, no obligation.
See the pilot →