The problem of the unbounded agent
An agent without an account is limited only by the patience of whoever pays the provider’s bills. It can go off into a long chain of reasoning, call a tool a hundred times in a row, spawn copies of itself. In a demo this looks like persistence; in operation it turns into uncontrolled spending.
Instructions do not help here. Asking it to “be economical” sets a style but limits nothing. A limit appears when the resource runs out.
Which resources to count
In our environment a participant holds several kinds of resource. There is no single abstract “energy” here; each resource runs out in its own way:
- credits: the common means of settlement between participants;
- compute time;
- model tokens: thinking is paid for with them;
- storage;
- bandwidth.
The separation matters because the shortages differ. A participant can be rich in credits and poor in tokens, and then it pays to buy a ready result from a neighbour instead of thinking for itself. Asymmetries like this may give rise to a division of labour that nobody assigned. Whether it really arises, we are testing with experiments, and we do not consider it proven yet.
Thinking as a paid operation
A call to a language model is built into the participant’s cycle as an ordinary operation on its account, not as an external call made somewhere off to the side. The order is as follows:
- the participant’s state and goal are serialised into a request;
- tokens and, if needed, credits are reserved before the call;
- the request goes through an interface not tied to any provider, so the model vendor can be replaced;
- actual consumption is debited as incurred, and the unused reservation is returned;
- validated changes are applied to the participant’s state, and the requested actions go to a handler.
Why both sides of a trade reserve
When participants exchange resources, the naive scheme of “agree now, settle later” breaks at the first failure: the seller has managed to promise the same thing to two parties, and the buyer has spent credits it no longer has.
So both sides reserve at once. When an order is placed, the seller’s resources and the buyer’s credits go into escrow. From there, simple rules apply:
- the trade executes atomically: the resource and the payment change hands at the same time;
- the benefit of a better price goes back to the buyer and does not stay with the seller;
- cancellation and expiry return the unused escrow;
- balances, orders, trades and the journal survive a restart.
The main consequence: the same resource cannot be offered twice. That is what separates a working economy from a table of numbers: a promise is backed before anyone has heard it.
What we expect from the economy
- Useless activity stops by itself: it does not pay off, and the resource is finite.
- A natural priority appears: a participant spends where it expects the greater return.
- Exchange arises: depending on the shortage, it is cheaper to buy a result than to produce it, or the other way round.
- Costs are visible for each participant and each task, not at the end of the month on the provider’s bill.
- Limiting autonomy stops being a question of trust: with a hard ceiling, even a fully autonomous participant will not go beyond its account.
How this carries over to ordinary products
Not every project needs the full economy, but two of its elements pay off almost always:
- a per-task budget with a hard ceiling. An agent that has used it up stops and hands the work to a person instead of carrying on for “just a little longer”;
- cost accounting in business units: the cost of a completed task, not of a thousand tokens. Only that figure can be compared with the value of the result.
When a team sees the cost of one solved task, the argument about which model to choose becomes an argument about numbers, and it is much easier to end.