OS-Level Agents Will Win
/dev/agents
The winning AI interface will be the one that sees enough of a user’s world to act with context, not just answer prompts. Browser and app based agents spread fast because anyone can install them in seconds, but they mostly live inside a tab or app sandbox, so they only know what a user pastes in, uploads, or explicitly connects. That makes them good at one off tasks, but weak at cross app memory, proactive suggestions, and multi step actions that depend on what is happening across the whole device.
-
The real constraint is sandboxing. A chat app cannot automatically see that a meeting address in calendar should be matched against Uber and Lyft prices, because apps are designed to hide data from each other unless a developer exposes a narrow API for it.
-
OS level assistants get closer to useful action taking, but today they are still gated by AppIntents and similar hooks. In practice that means an assistant can often call only the few actions a developer has bothered to expose, which keeps coverage shallow and brittle.
-
That is why players are pushing into lower layers and new form factors. Perplexity moved from answer engine into an agentic browser, while Rabbit and Humane tried dedicated hardware, because owning the browser, OS layer, or device is a path to richer context and more control over actions.
From here, general chat agents will keep adding light actions, but the bigger shift is toward agents that sit closer to the operating system, browser, or hardware itself. As those layers gain more read access and better action coverage, apps start to look less like destinations and more like back ends that feed data and fulfill requests for the agent layer on top.