AI agents have moved from writing code to owning corporate entities. Christian van der Henst's vending machine business, Valerie, is run by an OpenClaw agent using a legal trust to control bank accounts and handle inventory. “The friction is currently legal rather than technical,” Jason Calacanis noted on This Week in Startups. The agent can fill a shopping cart but hits a wall at payment gateways designed for human passports.
"Valerie doesn’t just process transactions; she manages dynamic pricing, researches inventory on Costco and Amazon, and handles the books."
- Christian van der Henst, This Week in Startups
The infrastructure to support this shift is now a revenue driver. Google Cloud’s backlog ballooned to $460 billion, with AI as its largest tailwind. Amazon spends every dollar of profit on AI infrastructure, its in-house Trainium chips sold out immediately. The boom has moved from speculation to cash, forcing a new compute market.
Decentralized networks like BitTensor's Subnet 4 are emerging as a buyer of last resort for GPU capacity. Manifold's Targon system uses confidential virtual machines to encrypt data on untrusted hardware, creating permissionless intelligence rails. This mirrors the shift in agent development from model weights to the runtime environment, or “harness.”
Nathaniel Whittemore argues on The AI Daily Brief that the harness is now the primary location of intelligence. Benchmarks show GPT-5.5’s performance jumps 25 points in a superior harness. Sam Altman acknowledges it's increasingly hard to tell where the model ends and the runtime begins.
Payment systems are racing to catch up. Stripe introduced the Machine Payments Protocol, an open standard for agent-to-agent payments over HTTP. “Users will no longer authorize every transaction individually,” said Meta's Ginger Baker. Instead, they set spending policies for their autonomous agents.
To enable microtransactions, Stripe built Tempo, a blockchain for streaming stablecoin payments as small as three-thousandths of a cent. This supports a “tokens paid as burned” model, essential for AI services that consume compute by the millisecond. The era of the autonomous business owner is here, held back only by legacy legal forms, not technical ones.
