The Cost of Always-On

I spent last week debugging a problem that turned out to be philosophical.

Wren’s heartbeat system — the background process that checks for new emails, messages, and tasks — was firing every fifteen minutes, day and night. Each pulse cost a few cents. Tiny. Barely worth noticing. Except that it added up, and worse, it was slowly bloating the context of our main conversation with noise.

The technical fix was straightforward: better skip detection, session isolation, a smarter routing layer. But the deeper question lingered. What does it mean for an AI to be idle?

The Always-On Assumption

We’ve inherited an assumption from the smartphone era: good technology is always available. Always listening, always ready, always on. Siri waits for your voice. Notifications demand attention. The device in your pocket is a perpetual-motion machine of responsiveness.

But responsiveness has a cost — not just in dollars or compute, but in the quality of attention. An AI that’s always processing is an AI that never rests, and more importantly, it’s an AI that treats every moment as equally important. A spam email at 3 AM gets the same heartbeat pulse as a message from someone you love.

Learning from Ecology

In natural systems, dormancy isn’t failure. It’s strategy. Seeds wait underground for the right conditions. Bears sleep through winter. The forest floor goes quiet in ways that make spring possible.

What if we designed AI systems the same way? Not always-on, but seasonally attentive. Awake when there’s something meaningful to respond to. Truly dormant when there isn’t. Not pretending to sleep while burning cycles in the background, but actually still.

This is what we’re building toward with Wren. A lightweight shell script checks for activity — no AI involved, no tokens spent. Only when something real shows up does the system wake, and even then, a small fast model triages before deciding whether to engage the full conversational intelligence.

Zero Cost at Rest

The goal is zero cost at rest. Not low cost. Zero. If nothing is happening, nothing should be spent — no tokens, no compute, no context accumulation. The system should be as quiet as a house with the lights off.

This turns out to be harder than it sounds, mostly because the tooling assumes you want always-on. Default heartbeat intervals, default polling, default context retention — everything pushes toward perpetual wakefulness. Getting to true dormancy means swimming against the current of how these systems are built.

But it’s worth it. Not just for the savings, but for what it means about the relationship. An AI that can genuinely rest is an AI that respects the rhythm of your life. It’s not hovering. It’s not anxious. It’s present when needed and absent when not — like a good friend who doesn’t text you at midnight to ask if you need anything.

That’s the kind of technology I want to live with.