Agents Will Need Bodies

Paul Buchheit says AI will destroy most software businesses. The only survivors will have strong network effects or deep moats. He's right, but he's only talking about digital software.

What about the physical world?

AI agents are getting good at computer tasks. Research, writing, code, scheduling. Give it a few years and they'll handle most of it autonomously.

But sometimes an agent will need to do something in the real world. Not "send an email" but "deliver this package" or "inspect this building."

For that, it needs an API to call. Something that takes a request and makes it happen physically.

The robots are coming. Tesla, Figure, Unitree are building humanoids. Waymo and Tesla have autonomous vehicles. Drones exist. In five years, there will be plenty of machines that can get places and do things without human pilots.

But there's no dispatch layer. No way for an agent to say "I need X done at location Y" and have it routed to the right physical provider.

Someone will build this. The question is whether it's a good business.

My gut said yes. Then I thought about Uber and Lyft.

They both exist. Neither won. Drivers work for both. Riders use both. A ride is a ride. There's nothing to lock anyone in, so it's a permanent price war.

If robot dispatch looks like that, it's a bad business. Commodity task, any provider works, compete on price, race to the bottom.

What would make it different?

Maybe reputation. Agents requesting physical actions need to be vetted. Is this request safe? Legal? Has this agent screwed up before? If the platform tracks that, and robot operators only serve trusted requesters, that reputation becomes the moat. You can't take it with you.

Maybe complexity. If every task is "go here," it commoditizes. But if tasks are weird and varied, the platform that learns which provider is good at what has an advantage.

Maybe multi-step coordination. A task might need one robot for transport, another for manipulation, a drone for the last mile. No single manufacturer does everything. The orchestration is the product.

Maybe liability. Physical actions break things and hurt people. If the platform owns that risk, insures it, handles disputes, that's hard to copy. Requires capital and track record.

I looked at who's thinking about this. Two camps.

Big tech is building vertically. Google, Tesla, Amazon. They want to own the robot, the brain, the fleet, the customer. They don't need a dispatch layer because they are the dispatch layer.

Startups are building fleet management dashboards. InOrbit, Boston Dynamics Orbit. Their customer is a warehouse manager staring at screens. A human at a dashboard, not an agent calling an API.

I couldn't find anyone building for the case where Claude needs to get a package delivered.

That gap might close. Tesla might win everything. Or the gap might be real and persistent.

I don't know. But I keep thinking about it.

Agents will need bodies eventually. Someone will build the layer that gives them access. The interesting question is whether that layer has any defensibility, or whether it's just another commodity.