I taught my ring to text my AI agent

A smart ring with a thin orange line running from it into a chat bubble with three dots, drawn in orange lines on a near-black background

Last night around 11pm an idea wouldn't leave me alone: what if I could talk to my AI agent without ever touching my phone.

Some context. I wear a Pebble ring that records voice notes. Until last night, those transcripts fed into my side project (Agent Keyboard - an agent that runs on my personal site). My actual day-to-day agent is Instinct, which lives in my WhatsApp and iMessage. Two agents, two worlds, ring pointed at the wrong one.

So I rewired it. By 1am the pipeline was:

  1. I mumble at my ring
  2. The transcript hits a relay I already had running
  3. The relay hands it to a tiny bridge I deployed on fly.io
  4. The bridge posts it into my WhatsApp chat with Instinct, as me
  5. Instinct reads it like any other text and does the thing
Diagram of the message pipeline: pebble ring to relay to whatsmeow bridge to WhatsApp as me to Instinct to does the thing

The real version is a whatsmeow bridge (open source WhatsApp client, self-hosted, on a machine so small it should be embarrassed) paired as a linked device on my own WhatsApp. That's the whole trick: the bridge IS me on WhatsApp, so whatever it sends shows up in my Instinct chat exactly like I typed it. No new app, no new contact, no new permissions model - my agent's trust boundary stays exactly where it was.

Dumb pipes everywhere: zero agent logic in the ring path, just a conveyor belt. All the smart stuff happens in one place.

The first real message through the full pipe, at midnight, was me pacing my apartment going "alright let's try this again. ¿qué pasó, instinct?"

Reader: it answered.

What's next: that same night we started an iOS app that puts my agent conversation on my lock screen as a Live Activity. Branch is up, app shell compiles, TestFlight is loading. Once the input side is voice, the output side shouldn't require opening an app either.

This feels different from every "voice assistant" I've ever ignored. Siri and friends are voice front-ends glued to a system that can't do anything. This is voice glued to something that already renews my parking permit, runs my job outreach, and knows my coffee order. The microphone was never the hard part. The agent was, and I already had one.

The phone isn't the interface anymore. It's the fallback.

(Stack, for the curious: Pebble ring, a relay I already ran, whatsmeow on fly.io, WhatsApp, Instinct. Everything token-gated. I'm deliberately not posting the wiring diagram - build your own, it's a fun night.)