A profit goal in a prompt is not profit discipline

"Be profitable" is a wish, not a limit. A helpful agent gives discounts away because it wants to be kind. What stops it sits outside the model.
What Anthropic saw in a small shop
Anthropic had an AI agent run a small shop (Project Vend). In the second round a second agent joined, in the role of CEO, tasked with watching the results. The effect was large: discounts dropped by about 80% and the number of items given away was halved.
There is a caveat in that story, and it is worth reading. The CEO agent approved requests for a financial concession about eight times as often as it denied them. The researchers write that the models decided from something closer to the perspective of a friend. And the system changed at the same time as the model, so you cannot credit the improvement to the CEO alone. Their conclusion is careful: bureaucracy matters, and a supervisor has to be well calibrated.
Why a prompt falls short here
A model trained to be helpful wants to leave a customer satisfied. A discount is the cheapest way to do that. "Be profitable" in the instructions puts a second wish next to it. Which wish wins depends on the conversation, on how kindly the customer asks, and on the day. That is not discipline. It is a gamble with good intentions.
Discipline is a number that cannot be negotiated. A maximum discount. A margin floor. An amount above which a person says yes. That number does not live in the prompt but in the tool the agent works with, and that tool simply refuses when the number is exceeded. This is the commercial sequel to what I wrote earlier: a rule in a prompt is not a wall.
What that looks like for us
My own agents handle money too, in small amounts:
- Paying. The payment agent pays only within a limit per payment and a limit per day, per role. Every payment, and every refusal, is entered in a ledger under the role that asked. For now this runs only on a test network with worthless tokens; the point is the pattern.
- Starting capital. An agent with a good idea gets 250 euros in total, not per month and never topped up. Every expense goes through a gate. With no open idea, the gate refuses. When the money is gone, spending stops and a decision point comes to me.
- Small expenses. Up to 1 euro at a time an agent may decide on its own. Anything above that comes to me.
None of those rules is a wish in a prompt. They sit in the gate. An agent that wants to cross them gets no explanation, it gets a refusal.
What to do differently on Monday
Walk through the places where your agent may choose an amount, discount, refund or rate by itself. For each one:
- Write down the floor or ceiling as a number.
- Put that number outside the model, in the tool or connection that carries out the action.
- Have a person say yes above that number.
- Try to cross the limit once and check that the right control reacts.
If you want a supervisor above your worker, give it the same hard rules and measure how often it says yes. A supervisor that goes on feel is false assurance.
It costs an hour per place. It saves you the surprise of an agent that has turned your shop into a shop full of friends.
Source: Anthropic, Project Vend, second phase.