Two agents, one €800 budget
Each agent stayed under its limit. Together, they spent €1,200.
Imagine asking an assistant to buy a monitor and a chair for your home office. €800 total, including delivery and taxes.
It delegates. One subagent finds the monitor, another finds the chair. Each gets permission to spend up to €600, since either purchase might need most of the budget.
They both find something for €600. They both buy it.
Congratulations on your €1,200 home office.
Neither agent exceeded its individual limit. Both bought what you asked for. The mistake was giving each one access to money the other could also spend.
Splitting work between agents is already a practical design choice. Anthropic's research system, for example, uses a lead agent to delegate searches to subagents working in parallel. Searching doesn't give those agents permission to buy things, of course. But once we delegate actions as well as research, we need to decide what happens to the authority behind the job.
Here, the €800 is the purchasing budget. The bill for running the models is a separate expense, which I am choosing not to think about for now.
Both permissions look fine
Suppose our imaginary system gives each subagent a signed credential. It identifies the job, limits the kind of purchase and sets a €600 ceiling. The shop checks the credential before accepting the order.
Each child's ceiling is lower than the parent's €800. Each purchase fits its child's ceiling. If those are all the checks, both orders pass.
There is a useful concept called attenuation: taking a permission and making it narrower. Biscuit, for example, lets a holder add restrictions to a token without contacting its issuer. The holder cannot use that process to grant more rights.
That is useful for saying "this delegate may only buy monitors" or "this request must be under €600". Our example asks an additional question: how much have all the delegates spent together?
An unchanged credential cannot answer that by itself. The shop needs information about other purchases, or an allocation which already accounted for them. This is a limit of independent checks, not a claim that Biscuit applications cannot enforce budgets.
The two orders don't even need to happen at the same time. Without shared accounting, the chair shop could accept its order tomorrow and still know nothing about today's monitor purchase.
Put the total somewhere
One straightforward design gives the job a shared budget record. Before a purchase, trusted code checks what remains and reserves the required amount in one atomic operation. In plain terms: another purchase cannot slip between checking the balance and claiming its share.
The monitor order reserves €600. Available budget falls to €200. The chair order asks to reserve another €600 and gets refused. The assistant can find something cheaper or ask you to change the budget.
Reading a shared balance is not enough. If both callers read €800, both decide €600 fits, and both update it afterwards, we have recreated the same problem with a database involved. Very enterprise.
Databases already have tools for this. PostgreSQL documents how a concurrent update to the same row can wait and then re-evaluate its condition against the updated row. A conditional budget update can use that behaviour. The reservation record and the balance change also need to commit together.
The component placing the order must enforce the reservation. If the agent can use another tool with unrestricted purchasing credentials, the budget record is just a polite suggestion.
An accepted order turns the reservation into spending. A definite rejection releases its reservation. A timeout leaves it unresolved: the shop might have accepted the order and lost the response. Releasing that money immediately could let the next purchase spend it again. A retry also needs to refer to the same order and reservation.
The model saying "I think it failed" is not enough to settle that.
Or split the money first
We could allocate €500 to the monitor and €300 to the chair before either starts shopping. Each side could then enforce its own allowance locally, without consulting a shared balance for every purchase.
That works if the allocations are exclusive: issuing them removes the full €800 from what the parent can spend or allocate elsewhere. Each allowance still needs protection against being spent twice. Making two copies of a €500 token does not create two €500 allocations.
The trade-off is flexibility. A €550 monitor and a €250 chair fit the original budget, but require moving €50 between those shares. That transfer needs coordination too.
Wrapping up
For a small system, I would start with one place that authorizes purchases against the job's remaining budget. Subagents can research in parallel and bring their choices back. Giving every researcher its own purchasing authority is optional.
When I see "each agent has a spending limit", I want to know whether those limits draw from one allowance or create separate ones. For this job, every purchase needs to leave less of the same €800 for everybody else.