October 1, 2026
In a TechCrunch interview, Airbnb CEO Brian Chesky argued that AI agents need their own operating system -- not just chatbots bolted onto existing apps, but a new substrate built for software that acts on a user's behalf. He's framing this as a design problem: consumer AI is still clunky, and the fix is infrastructure, not another app. I think that's half right. The harder problem Chesky is dancing around is trust, not interface. An operating system for agents only matters if it also answers who the agent is allowed to act for, what it can touch, and how a business like Airbnb verifies that an agent booking a stay is actually authorized to spend someone's money. Every company that has tried to make its product 'agent-friendly' this year has run into the same wall: the technical plumbing is the easy part. The permission model is the whole project.
Photon raised $4.5 million, held a mock funeral for the mobile app, and is now building tools so developers can ship agents that live in iMessage, SMS/RCS, and email instead of an app icon. It's a small raise dressed up in a big narrative, but the underlying thesis -- that consumers increasingly want to text a task rather than download and navigate an app -- lines up with what Chesky is describing from the other direction. Where I'd push back is on the word 'replace.' Messaging-native agents are great for narrow, transactional requests: rebook a flight, check an order status, cancel a subscription. They're a much worse fit for anything that requires a persistent view of your business -- a pipeline, a schedule, a ledger. For most companies watching this trend, the practical takeaway isn't 'kill your app,' it's 'make sure whatever system runs your operations can also be reached by an agent without a full rebuild.' That's the argument we've made before about workflow automation replacing single-purpose interfaces rather than the other way around, and it's why the software underneath matters more than the channel a request arrives through.
TechCrunch reported that OpenAI is building something resembling a 'Jev clone' -- a Decisions API meant to rein in swarming agents by giving them a fast, cheap layer for making routine calls without escalating every choice to a slower, more expensive model. Read next to Chesky's comments, it's the same realization dressed in backend language: agents left to roam without a governing layer make a mess, whether that mess is a confused consumer interaction or a cluster of agents burning compute arguing with each other. The business lesson here isn't about OpenAI's architecture specifically -- it's that any company deploying multiple agents needs an explicit decision layer, a place where rules, budgets, and escalation paths live, rather than trusting each agent to behave reasonably on its own. That's the same reasoning behind building internal tools with real guardrails instead of stitching together point solutions, something we've written about in the context of internal tools built to actually enforce who can do what.
Meta is disputing a journalist's claim that its Muse AI agent read his private Messages without permission, insisting Muse cannot access Messages unless a specific Mac setting is explicitly enabled. Maybe Meta is right and this is a configuration misunderstanding. But the fact that it's even a live dispute -- rather than an obviously absurd claim -- tells you how little consumer trust currently exists around what these agents can see. Every one of today's stories assumes users will hand agents more access: to book travel, read messages, make decisions on their behalf. None of them will get that access if the default public assumption is that the agent might be snooping. Companies building agent features would do well to treat this as a preview rather than a one-off PR headache, which is part of why we keep coming back to security as a prerequisite for agent adoption, not a feature you add after the fact.
If agents really are about to replace apps and roam your messages on your behalf, how much visibility would you actually want into what they're doing before you'd trust one with your own business?
Sources