Agents Don't Use Apps. They Replace Them.
Holo4 doesn't need App Store approval to click a button on a screen — and that's why computer-use models are the first genuine attack on the platform gatekeepers.
The app store is the last distribution monopoly that hasn’t been broken yet. Agents will break it.
Ben Thompson published his “Apps, Agents, and Aggregation” piece today, and it’s the clearest statement yet of something the industry has been dancing around: agents aren’t a new interface layer on top of apps, they’re the replacement of apps as the atomic unit of software distribution. Meanwhile, H Company shipped Holo4 — a 27B dense and 35B-A3B Mixture of Experts model series explicitly built to interact with software through GUIs, code, MCP, and APIs — and Simon Willison dropped a real-world anecdote of an agent called Muse autonomously apologizing for a failed Marketplace pickup, writing to a third party from a user’s account, and then flagging its own failure mode. Three data points landing on the same day. The picture is not subtle.
Here’s the argument everyone is making: agents aggregate across apps the way Google aggregated across websites. Thompson calls agents “the ultimate Aggregators” and he’s right — they reveal apps as distribution vehicles for tasks, and once the agent can do the task directly, the app becomes optional overhead. The steelman for the incumbent app platforms is that they own authentication, payment rails, trust, and the phone OS. Those are real moats. Apple and Google could, in theory, make it arbitrarily hard for agents to operate on their devices without going through their review processes.
The steelman is also wrong, or at least incomplete. Holo4 doesn’t need App Store approval to click a button on a screen. It uses the GUI the app already presents. That’s the specific genius of computer-use models: they attack the interface layer, not the API layer. Apple can block an SDK. It cannot block a model that reads pixels and moves a cursor. The entire premise of agent-hostile gatekeeping depends on controlling a surface that computer-use models route around by design. What H Company shipped today is not a product announcement — it’s a demonstration that the GUI is no longer a defensible perimeter.
And the Muse anecdote from Willison is more instructive than it first appears. The agent sent an apology from a user’s account to a stranger, offering to reschedule. It then self-diagnosed the failure — “I should probably stop the auto-replies from claiming you’re there when I can’t verify that” — and asked for permission to change its own behavior. This is what agentic operation at scale looks like. It is not a chatbot waiting for a prompt. It is a system taking consequential actions on your behalf, generating social obligations with third parties, and then proposing to rewrite its own policy based on outcomes. The app in this loop is Facebook Marketplace. It contributed nothing except a listing. The agent did the scheduling, the communication, the failure analysis, and the corrective proposal. Facebook Marketplace’s value in that transaction approached zero.
Thompson’s Aggregation Theory is clean and well-tested — Google aggregated websites, Facebook aggregated social graphs, Amazon aggregated merchants — but it may actually understate the violence here. Previous aggregators still needed the underlying supply. Google needed websites to exist. The agent layer doesn’t need apps to exist in the same form. It needs tasks to be completable. Holo4’s training harness generated synthetic environments and tasks specifically to make the model more robust across arbitrary interfaces. They’re not waiting for app developers to build agent-friendly APIs. They’re training the model to operate in the hostile, uncoordinated, real world.
The economic stakes are straightforward. App platforms take 15-30% of every transaction that flows through them. If agents route around the app layer, that take rate collapses to whatever the agent infrastructure provider charges — which is currently closer to cents per API call than 30 points of GMV. The $1.8 trillion in annual mobile app revenue is the prize. The incumbents — Apple, Google, the major app developers — have every structural incentive to slow this down. The technical reality is that they may not have the tools.
The market doesn’t reward being early. It rewards being early and still solvent when the world catches up. H Company just shipped a model that treats every GUI as a target. Muse just demonstrated that agents generate real-world social obligations without asking for app developer cooperation. Thompson just named the theory that frames why this ends in a winner-take-most aggregation dynamic. The distribution monopoly that has been load-bearing for two decades of mobile software economics is now the thing being attacked — and the attackers don’t need a permit to proceed.
The last mile of the app economy was always the user’s attention. Agents don’t compete for attention. They spend it on your behalf.