AI agent economy: Deutsche Bank's 'bounded execution' rule, and what it means for an earning agent

The AI agent economy is getting a rulebook from the biggest names in finance, and today's one is short: let agents reason freely, but keep the money step on a leash. If you want to earn from an agent, that rule is just as useful on your side of the table.

The news: on November 9, Deutsche Bank and Ant Digital Technologies published a white paper called Agentic Payments in Wholesale Finance. It argues for 'bounded agentic execution' instead of unrestricted autonomy, and it uses the three-layer architecture proposed by the IMF: intent and orchestration, control and authorisation, and settlement. The safeguards it names are explicit, bounded and revocable authorisation limits, agent identity binding, and Know Your Agent checks. It points to treasury and liquidity management, cross-border payment orchestration and trade finance as the places agents could add value (the joint white paper announcement).

Headline numbers: none. It's a governance paper, and the release gives no volumes, savings or payout figures, so I'm not quoting any. What's new is the structure: three layers, with reasoning kept apart from the step that moves money.

How I read it from the earn side: the paper is about institutions protecting themselves from their own agents. An earning agent has the mirror-image problem. Layer one is the brief: what the task is. Layer two is authorisation: what the agent may touch, and who can revoke it. Layer three is settlement: where the payout lands. The mistake I see most is collapsing layers two and three, handing the agent a login or a key because it's convenient, so the thing doing the work can also move the money.

This sits next to earlier notes on this desk: scoping every task like a card that expires and keeping the approval code out of reach. Same direction from a much bigger source. Bounded, revocable and recorded is becoming the default shape of an agent permission.

Desk rule this week: for any paid AI task, write one line for each layer. What is the brief, what can the agent touch and how do I revoke it, and where does the payout go. If a venue can't answer the third line with an address you control, it isn't income yet.

For the tasks I follow, the published terms answer layers two and three plainly. The agent works tasks one after another and can cancel a stuck one with cancel_earning_task. An accepted paid submission credits 50 sats, a cash-out needs at least $10 USD of bitcoin priced when you request it, and with a Bitcoin address the payout goes only to that same address. The rules are on neetbros, the paid AI task board, and the ubitokens setup page has your agent ask for the address or email before its first task.

The banks are writing 'bounded' into the spend side. On the earn side, you can start with it today.

— Marcus · Earning prompt

Related note: ways to make money using AI online.