Voice notifications for AI coding agents

The problem with visual monitoring

You're running three AI coding agents in parallel. One is refactoring auth, another is writing tests, a third is fixing a deploy script. Each one might finish, fail, or get stuck at any moment. How do you know what's happening?

The common answers all share a flaw: they require your eyes on a screen.

Terminal grids work if you're sitting at your desk staring at them. Walk to the kitchen and you're blind. Slack webhooks flood a channel with machine-generated noise that buries your real conversations. Desktop notifications stack into an unreadable pile when five events arrive in thirty seconds. All of these assume you're watching something. None of them work when you're in another room, making coffee, or focused on a different task.

The gap is audio. There is no sound. Your agents could be on fire and you wouldn't hear a thing.

What voice notifications look like in practice

Radio Agent fills that gap. It runs a radio station on your local network that plays ambient music and speaks announcements when your agents do something worth knowing about.

Here's what you actually hear during a typical session:

Background music plays. Quiet, ambient. You stop noticing it after a few minutes.

An agent finishes a task. The music dips, a voice says "Config module shipped with tests. Clean commit." The music comes back. Five seconds total. You heard what happened from the next room without looking at anything.

A build fails. The music dips, and a different, male voice says "Build failed for the auth module. Dependency issue." The voice change tells you something went wrong before the words even register. You walk back to your desk knowing exactly what needs attention.

An agent picks up new work. No voice this time. Just a soft rising chime under the music. You don't consciously notice it, but you know the session is active.

Three audio channels, one stream

Radio Agent mixes three independent layers into a single audio stream that any device on your network can play:

Channel	What it carries	How it feels
Music	Continuous ambient tracks from your library	Background texture. You stop hearing it.
Voice	Spoken announcements via Kokoro TTS	Pulls attention briefly, then disappears.
Tones	Short sound effects (chime, chord, pulse)	Subliminal. You feel the session rhythm.

Each channel can be muted independently. Keep voice and mute music. Keep tones and mute voice. Whatever combination fits the moment.

How it connects to your agents

Any system that can make an HTTP POST can send announcements. The webhook API accepts JSON with one required field:

curl -X POST http://your-host:8001/announce \
  -H 'Content-Type: application/json' \
  -d '{"detail": "Auth refactor complete", "kind": "agent.completed", "agent": "eng1"}'

The kind field controls routing. Completions get a warm voice and a resolved chord. Failures get a different voice and a tense tone. Starts and stops play a chime with no voice, keeping the stream from getting chatty.

If you use Claude Code, the setup is even simpler: install the DJ skill and the setup skill, point them at your Radio Agent instance, and every agent announcement gets rewritten into natural radio-style speech before broadcasting.

Why audio instead of more visual tools

	Visual tools	Audio (Radio Agent)
Away from desk	You miss everything	You hear it from any room
Multi-agent scale	N windows or channels	One stream, mixed
Attention cost	Context switch to read	Peripheral, no switch
Failure detection	Scan for red badges	Different voice, instant
Setup	Per-tool integration	One webhook endpoint

This isn't an argument against dashboards or Slack. Those are the right tools for async discussion, audit trails, and detailed investigation. Radio Agent handles a different job: real-time ambient awareness. Knowing the state of your agents the way you know whether it's raining outside. Not because you checked, but because you can hear it.