Picture an AI agent tasked with monitoring a factory line, checking inventory every few minutes, or quietly organizing files on a laptop all day. In theory, this is exactly what agentic AI promises: systems that work in the background without a human hovering over them. In practice, almost none of this happens today because true “always on” agents aren’t scalable.
Most AI agents run on cloud APIs, and cloud APIs charge per token. An agent that checks something every five minutes, all day, every day, racks up a bill that scales with how often it works, not with how useful that work is. So companies ration their agents. They run them on a schedule instead of continuously. They cap how many tasks an agent can attempt before someone has to approve the next round of API spend. The agent that could, technically, monitor everything in real time instead gets throttled down to checking in periodically, because someone has to pay for every token it thinks with.
We already ration compute in ways that would look absurd if applied to a human employee. Nobody tells a warehouse worker to check inventory once every four hours because looking costs money. Nobody tells a security guard to only glance at the monitors every thirty minutes to save on electricity. Attention, for a person, is a fixed cost of showing up to the job. For an AI agent, attention is a metered expense, and every additional glance shows up on an invoice.
Where the Current Model Runs Out of Room
Today, most AI agents send every request to the cloud, wait for a response, and do it all over again the next time they need to think. A device or application sends a prompt to a hosted model, waits, and pays for the privilege each time. That's a fine deal for occasional requests, like asking a chatbot to draft an email. It's a bad deal for the kind of work that's supposed to define the next phase of agentic AI: background tasks, ambient monitoring, agents that live inside a device rather than phone home to one. If checking in costs money every time, you build agents that check in as rarely as possible.
But what if an AI model was set up to bypass the cloud, with fixed costs that aren’t based on usage? If checking in is basically free because the model already lives on the device, you build agents that check in all the time, because there's no longer a reason to hold back. That difference changes what companies decide to build, not just how much it costs them. Instead of asking an agent to summarize a report once a day, it can watch for changes continuously. Instead of periodically checking on machine health, it can check for errors as they happen. Instead of waiting for someone to ask for help, it can quietly have the answer ready before the question is asked.
Pricing models don't just determine what AI costs. They determine what AI gets built at all. A lot of useful agents never got built simply because nobody could justify the bill for running them all the time.
Why This Is Changing Now
Only recently have small models become capable enough that moving them onto a device no longer means giving up most of what makes an agent useful. These models can now run locally on consumer devices while still handling the kinds of multi-step tasks that previously required much larger cloud-hosted systems.
Once inference happens on the device itself, the per-token bill disappears, and the marginal cost of checking again becomes negligible compared with paying a cloud API every time. There's still a real cost, in battery, compute cycles, and heat, but it's a cost the device already absorbs by existing, not a line item that grows every time the agent looks again. That's the difference that makes background monitoring, ambient assistants, and agents embedded in appliances, cars, and industrial equipment plausible at scale, not as a demo, but as something a company would actually leave running.
Every time AI becomes cheap enough to run continuously, it stops behaving like software you open and starts behaving like infrastructure you simply expect to be there.
None of this means cloud-scale models are obsolete. Plenty of tasks will keep justifying a data center round trip, and larger models still hold real advantages on harder problems. But the biggest shift underway may not be from larger models to smaller ones. It may be from AI that's consulted occasionally to AI that's continuously present.
For years, we've measured AI by how intelligent it could become. The next phase may be measured by something simpler: whether that intelligence can afford to stay awake.
