Real life is multiplayer

In the same week, Instinct and Meta's Muse announced that their AI agents could make phone calls. The headlines treated it as another round of feature race. What caught my attention was not the call itself, but the reason both products had arrived there.

The moment software is expected to finish a task, it runs into the parts of the world that software never fully absorbed.

The dentist keeps a cancellation list by phone. The restarurant does not take online reservations. The insurance company needs 3 transfers before someone can explain the bill. The repair shop wants to know what the strange noise sounds like. Behind every clean digital workflow is an exception that eventually reaches a person.

For years, we designed software around the assumption that the world would become structure enough for machines. Every business would expose an API. Every service would have an online booking system. Every transaction would fit neatly in to a form.

The world did not cooperate. Instead AI is learning to operate inside the mess.

Consider a simple request: book a pediatrician appointment. The phone call typically lasts only two minutes, but completing the task requires much more than speech. The agent needs to know which child needs the appointment, which doctor family prefers, what the insurance covers, when both parents are available and whether pulling the child out of school is reasonable.

Then the receptionist offers a few imperfect options: another doctor tomorrow, the preferred doctor next month or a same-day appointment at a far off location.

Can the agent choose; can it move another commitment? Can it share the child's DoB? Can it pay a deposit? At what point should it stop to bring a parent into the decision?

Making the call is a capability. Carrying enough context, permission and judgement to complete the task is the product.

I first encountered this problem at Google, where I helped start Assistant for Families. We kept finding that a household could not be modeled as one user with a longer contact list. A family is a living system of shared responsibilities, competing preferences and decisions that ripple across several people.

The models have changed enormously since then. Family life has not.

At Ollie AI, we begin with that same reality. The work may move through a browser, an email, a calendar, a text message or eventually a phone call. Those are simply different routes through the existing system. A family should not need to know which route the assistant took. They care that the appointment was booked, the bill was fixed and the reservation was made without the logistics returning to their plate.

Phone calls will quickly become standard across capable agents. Instinct and Muse shipping them at nearly identical times shows how quickly individual features reach parity. The lasting difference will come from everything surrounding the call; knowing whom to contact, carrying the accurate context into the conversation and recognizing when the agent no longer has the permission to decide.

This is the larger transition that's underway. AI is leaving single player world and entering the shared, unfinished world where people actually live.

Intelligence alone will not be enough there. These products will need memory without surveillance, initiative without the recklessness and utility that earns trust over time. They will need to coordinate across people who do not always agree. They will need to act without making their users feel that they have surrendered control.

There are easier places to build AI than inside the disorder of everyday life. There may be none where getting it right matters more.