Factories used to generate their own power. Every mill had its boiler room, its engineers, its coal deliveries, because power was a thing you made, not a thing you bought. Then the grid arrived, and within a generation the boiler room was gone and nobody mourned it. The companies that won weren't the ones with the best boilers. They were the ones that redirected all that attention to what they actually made.
Intelligence is going the same way, and a team of AI agents is the clearest place to see it. Ours run on Claude. We didn't train anything. That's not a compromise. It's the design.
The frontier is a utility now
The models improving fastest in the world are built by a handful of labs spending sums no operating company can match, and they're available over a wire. When the frontier moves, every agent on our roster gets smarter without a single one of them changing jobs. The writer, the fact-checker, the researchers, all of them inherited every capability jump the model made, the way every factory on the grid inherited every improvement at the power station.
Owning the model was never the differentiating asset for a company like ours, any more than owning a turbine was for a textile mill. Renting the frontier means the intelligence is always current, and always someone else's full-time job. It also means you can pick the best model for each job, and we test the models on real business work to know which one that is.
When the model gets smarter, the whole team gets smarter overnight, and nobody had to be retrained, replaced, or migrated. That's what building on the frontier buys.
What an agent adds that a model isn't
If intelligence is rented, what's actually ours? Everything that makes intelligence employable. A model is capability in a jar: brilliant, generic, and amnesiac. An agent is that capability given a role, a memory of you and the company, an address, standing routines, a manager, and a place where its work lands. Same underlying mind; entirely different economic object. The difference between Claude and an agent running on Claude is the difference between electricity and a factory.
This division of labor is also why the whole arrangement stays trustworthy. The lab makes the model capable and safe. The team structure makes it accountable: scoped to a job, visible in its work, answerable to someone. Neither layer can do the other's job.
One more thing renting gets you: your own keys
There's a version of this architecture where the agent company sits between you and your tools, holding your credentials. Ours runs the other way: your agents run on your Mac or in the cloud, your choice, and the accounts and tools you connect stay yours. Where your agents run turns out to matter as much as what they run on.
The boiler-room era of AI, every company training its own, ended before most companies noticed it had started. What's left is the interesting work: not making intelligence, but hiring it, structuring it, and pointing it at something worth doing — which is exactly the work waiting where your agents already are.
