Every notification is an interrupt. Every interrupt costs you somewhere between a few seconds and twenty-five minutes, depending on how deep you were in the work when it arrived. If you run AI coding agents, the interrupt math gets worse fast. Three agents generating five events each per hour means fifteen context switches if you're paying attention to all of them, or zero awareness if you're ignoring the channel they post to.
There's a middle ground that the notification industry mostly ignores: information that reaches you without demanding your attention. Information you become aware of without choosing to look at it. The academic term is peripheral awareness. The practical term is calm technology.
The attention spectrum
Not all information deserves the same urgency. A completed task and a failed build are both events, but they carry different weight. The problem with most notification systems is that they flatten this distinction. A Slack message is a Slack message whether it says "tests passed" or "production is down." A desktop toast is a desktop toast. Everything arrives at the same volume, in the same visual format, competing for the same slice of attention.
A better model sorts information into three tiers based on how urgently you need to act on it:
- Peripheral: you should be aware of it, but you don't need to act. An agent starting a task. A test suite passing. Music playing.
- Informational: you should know about it soon, but it can wait until a natural break. A feature was deployed. A code review is ready. A voice speaking over the music.
- Interruptive: you need to act now. A build is blocking other agents. A deploy failed. A phone buzzing on your wrist.
Each tier should use a different channel. Peripheral information should not buzz your phone. Interruptive information should not sit quietly in a feed you check twice a day.
Three layers for a multi-agent workflow
Here's a concrete setup that implements this model:
The layers are not redundant. They serve different situations. When you're at your desk, the dashboard gives you detail the audio can't. When you're in the kitchen, the audio gives you awareness the dashboard can't. When you're outside, the push notification gives you urgency the audio can't.
The critical design choice: each event is routed to exactly the layer that matches its urgency. An agent starting a task does not buzz your watch. A blocked deploy does not sit quietly in a stream you might not be listening to.
What calm technology actually means
Mark Weiser and John Seely Brown coined the term in 1996. Their central idea was simple: the best technology moves between the periphery and center of your attention fluidly, and most of the time it stays in the periphery.
Calm technology engages both the center and the periphery of our attention, and in fact moves back and forth between the two. Weiser & Brown, "The Coming Age of Calm Technology," 1996
Their example was a dangling string connected to a motor that responds to network traffic. The string whirls when the network is busy, twitches when it's quiet. Nobody watches the string. Everyone in the office has a sense of whether the network is healthy. The string lives in the periphery until something unusual happens, and then it pulls your attention to the center for a moment.
Audio is a natural fit for this pattern. Music sits in the periphery. You stop hearing it after a few minutes, the way you stop noticing the hum of an air conditioner. When a voice speaks over the music, it pulls information to the center of your attention for five seconds, then the music returns and attention moves back to the periphery.
Compare this to a notification badge. The badge sits at the center of your attention from the moment it appears until the moment you dismiss it. It doesn't move to the periphery. It just sits there, accumulating a count. That's the opposite of calm.
How Radio Agent implements this
Radio Agent is an open-source tool that turns AI agent activity into an ambient audio stream. It implements the three-layer model described above, with audio as the primary awareness channel.
The audio stream mixes three channels:
- Music plays continuously. Ambient tracks from your library, shuffled with crossfades. This is the peripheral bed. It tells you the system is alive without saying anything specific.
- Voice speaks announcements when agents complete tasks, fail, or get stuck. The music dips, the voice speaks for five seconds, the music returns. This is the center-pull moment. Failures use a different voice than completions, so your ear differentiates before your brain parses the words.
- Tones are short sound effects (a rising chime, a resolved chord, a pulse) that play under the music at low volume. They signal frequent events like "agent started" or "agent idle" without voice, because speaking those aloud every thirty seconds would be exhausting. After about thirty minutes, you stop consciously hearing the tones but you develop a sense of whether the session is active or quiet.
Each channel maps to a different position on the attention spectrum. Music is fully peripheral. Tones are near-peripheral. Voice is briefly central. Push notifications (configured separately) are fully interruptive.
The practical test
The question isn't whether this architecture is theoretically sound. The question is whether it actually works when you're making coffee and three agents are running.
Here's the test: after an hour with Radio Agent playing in the background, can you answer "what's the general state of the session?" without having looked at a screen? If you can say "two agents are active, one finished a while ago, nothing has failed" based purely on what you heard, the ambient layer is working. If you can't, something is wrong with the event routing, the volume balance, or the tone design.
The second test: were you ever surprised by something you should have known about? If a build failed ten minutes ago and you didn't know until you walked back to your desk, the audio layer failed to surface it. That's a bug in the notification routing, and fixing it usually means adjusting which events get voice announcements versus tone-only signals.
Both tests are subjective, which is the point. Ambient awareness is a felt experience, not a metric. You know it's working when you stop thinking about it and just know.