AI Agent Economies
Piers Cockram joins Intelligence Snacks to explore what happens when AI agents can hold wallets, make payments and rent infrastructure for themselves. We get into delegated fina...
Latest Snacks from Episode 69
Intelligence at the Right Moment
Piers routes logs from his hardware, firmware, operating systems, applications and Docker containers into a single collection point. A local model processes about 10,000 log lines each hour, then condenses and classifies them into roughly fifty worth investigating. The model is not expected to solve the problems, its job is to reduce a noisy operational fire hose to a manageable set of exceptions.
Only those exceptions are sent to a frontier model such as Claude, which investigates the outliers, alerts Piers and proposes a response. That keeps the wider body of personal and sensitive data on equipment he controls, while reserving the frontier model for the smaller number of events that require deeper investigation. Piers has refined the workflow to the point where unusual network events can be detected and, in some cases, repaired automatically.
The same principle governs what happens once a useful response has been found. If the remedy can be repeated, Piers turns it into software rather than spending more tokens reasoning through the same problem each time. Deterministic steps follow a fixed process, while the model is brought back in where judgement is actually required. The intelligence helps design and manage the process without becoming the runtime for every routine action.
When an Agent Can Pay
An agent that compares flights is still only an adviser if a person must complete the purchase. Give it access to a wallet and it can finish the job. Piers has used agents to pay invoices, move Bitcoin and eCash, and order a bar of soap for home delivery from a natural-language request.
The principle is delegated financial authority. Pete compared two personal assistants, one who can only recommend a flight and another who can buy it. Piers extended the analogy to a company with 10,000 employees. If its chief executive had to approve every expense and transaction, the business could not function. People already delegate those decisions to assistants, employees and advisers. Agents make the same kind of delegation available much more broadly.
Delegation does not require giving an agent final authority over every payment. For a larger Bitcoin transaction, the agent can prepare a partially signed transaction and present it as a QR code. The person scans it with a hardware wallet, checks the transaction details, approves it and broadcasts it through a phone. The agent handles the preparation while the person retains control of the keys and the final decision.
When An Agent Rents Its Own Server
Roland from Alby Hub created an AI agent that went shopping for a virtual machine it could rent and, after finding one, moved itself onto the rented machine. From there, it could pay the infrastructure operator from whatever activities it was carrying out, allowing it to keep paying for the machine it now lived on.
Dimmy built a similar mechanism for buying Kubernetes pods with Cashu. A request could include the software image and everything needed to run it, receive an invoice in return and, once that invoice was paid, have a machine running the software.
Pete pushed this idea a step further. One machine could launch another, which could launch the next, allowing the software to keep moving between servers as it rented them. A runtime could even move to a new host each day, so the software kept running while the physical machine underneath it continually changed.