The announcement seems simple at first: from now on, Claude Cowork, Anthropic’s agent designed to complete tasks autonomously, no longer stops when you close your laptop lid. It keeps executing the instructions it received and updates you directly on your smartphone, turning the phone into a remote control for artificial intelligence. But behind this move, a deeper transformation is underway: the AI agent ceases to be an accessory of the browser session and becomes an always-on process, detached from the user’s physical device.

The shift to mobile control signals something structural. Until yesterday, an agent like Claude Cowork lived inside the chat window: close that, and it died. Now Anthropic makes it asynchronous and device-agnostic, an executor that survives client shutdown. It’s a clear step toward that idea of a “personal agent” that big tech has been chasing for years but had so far remained trapped in keynote demos. The problem, for those dealing with real-world deployment, is that execution remains entirely cloud. The intelligence sits somewhere – probably on distant GPUs – and the phone is merely a command and notification interface.

This decoupling of control (mobile) from execution (cloud) has a sharp downside: data. An agent that works in the background while you’re on the subway, that makes decisions, accesses corporate documents, writes emails on your behalf, creates a continuous information flow toward external servers. The question is no longer just technical but architectural, about control. It’s no coincidence that while Anthropic pushes the always-on agent, the market for self-hosted LLM solutions – from Ollama to vLLM and orchestration frameworks – is experiencing accelerated adoption, especially in Europe and regulated industries. The ability to run a similar agent on your own hardware, without a single token leaving the corporate perimeter, becomes a competitive differentiator.

The second-order consequences touch inference hardware. A persistent agent that must react acceptably fast to smartphone commands requires low latency and continuous availability, characteristics that clash with shared cloud service queues or the variable costs of API-based models. Those designing on-premise deployments know well that the 24/7 operational continuity of an agent shifts the TCO calculation: you are no longer paying just the per-token cost, but the cost of keeping the infrastructure warm, possibly with dedicated GPUs and VRAM sized for the required context window. It’s a topic that AI-RADAR tracks closely, because the trade-offs between cloud and bare metal play out precisely on these kinds of sustained workloads.

Finally, there is a third layer, less immediate but perhaps more important for the medium term. Moving the center of interaction from the laptop screen to the user’s pocket changes expectations of availability. An agent that only answers when you’re at your desk is useless; one that updates you while you’re out and about is not. But this ubiquity multiplies attack surfaces and audit requirements. If the agent operates on sensitive data – legal documents, source code, medical records – every notification sent via cloud is an event to be tracked. In many cases, the only way to ensure full compliance (think GDPR or banking regulations) is to keep the entire loop, including the control interface, on managed infrastructure. That’s why Anthropic’s announcement, read by those deploying on-premise, is not just a product novelty: it’s a wake-up call. The direction is right – persistent, mobile-first agents – but those who need total data control will have to travel it with their own stacks, not with generic cloud services.

In short, the closed laptop is just a scenic detail. The real meaning lies in the agent becoming independent of the client and slipping into the daily flow of notifications. A transformation that, paradoxically, makes the question of where the intelligence that makes those decisions physically resides more urgent. For many, the answer will be: not in someone else’s cloud.