May 2026

I built an agent that runs my WhatsApp

I was missing important stuff in the groups every week. So I put an agent to read my WhatsApp for me, send me what needs an urgent reply at the end of the day, and a summary of every contact and every group on Sunday. Runs on the home PC, costs zero, plugged into my own number.

AI agent reading my WhatsApp messages in the background

I was spending way too much time on WhatsApp. Supplier groups, forwarding stuff to the manager, replying to quotes, the family group I had already given up on two days earlier.

Every night, the same feeling: something important slipped through and I have no idea what.

Sometimes it was paranoia. Sometimes it wasn’t. I’d find out on Monday when a supplier was chasing a reply that never came.

Instead of a bot that talks, a bot that reads

I was already building My Digital Cortex. Memory ready, persona ready. What was missing was the part I called “agency”: the agent stops just answering and starts acting in my world.

WhatsApp was the most obvious place to start. That’s where my attention leaks first.

Instead of a bot that talks to customers, a bot that reads me.

What it does, in three lines:

Whatever it observes flows back into the Cortex. Every other AI tool I use afterwards already knows who I’ve been talking to, what about, and what’s still pending.

Built it in one night

Runs on the PC that’s already on at home. Plugged into my own number. Sees everything I see, and talks to nobody.

The first week paid off. Sunday night the summary hit my phone and there were two things I hadn’t seen:

  1. A supplier chasing a quote I swore I had answered.
  2. A reminder from my kid’s school. I was about to miss the deadline.

Neither was going to break me. Both were going to give me a headache.

The problem wasn’t WhatsApp. It was attention.

The part that got me the most

It wasn’t the group summaries. It was the per-contact summaries.

For each relevant person, it:

I used to remember people in a panic, only when something was already chasing me. Now there’s something pushing it up to me every week, before it turns into a problem.

Bonus: Sunday night it reads my own messages from the week and updates my profile inside the Cortex. No invention, just what can be observed from what I actually wrote.

The invention trap

This is exactly where a loose model invents the universe.

The first version handed me a novel. It drew up a psychological profile of people I had traded three messages with. It deduced intentions nobody had stated. It looked deep. It was made up.

I rewrote the prompt forcing a fixed format and forbidding invention. If there’s nothing to say, say nothing.

I’d rather have nothing than wrong.

This bot writes into the Cortex. If it makes things up, next week I’m making decisions based on its made-up nonsense. Not worth it.

Plan B, C, D

To avoid depending on a single model, I built in redundancy. If the first one is down or out of quota, it tries the second. If the second is down, it goes to the third.

It’s not a luxury:

The bot has to run. Cost so far: zero. Everything fits inside free tiers at the scale of one person.

Phase 2: letting it reply

Phase 1 is reading. Phase 2 is letting it reply to small stuff for me.

“On my way”. “Got it”. “What time did we agree on”. Confirming a payment finance already passed me.

No end customer. No decision. Just the thank-yous and the confirmations that eat my day without adding anything.

For that to work without turning into a nightmare, I need three things:

  1. Knowing who each contact is. Phase 1 is already building that inside the Cortex.
  2. Clear rules on what it answers alone, what it drafts for me to approve with one click, and what it never touches.
  3. Learning the way I write, so its replies sound like me.

All three come from the same source: the agent reading what I myself write.

For now it only reads. And just from reading, it has already changed the way I show up on Monday morning.

Stack