Delegation is the whole point of an agent: you hand it authority so it can act without you. That only works if the authority is bounded and the boundaries are clear. In payments, a well-governed mandate is not a vague permission to spend — it is a precise statement of what the agent may do and where it must stop. Three controls do most of the work.

The three controls

  • Limits. How much, in total or per operation, the agent may move. Limits turn "spend responsibly" into a number.
  • Approved counterparties. Who the agent may pay. An allowlist prevents an agent from inventing a recipient.
  • Approval thresholds. Above what amount a human must approve. Thresholds reserve judgement for the consequential.

Together they define the shape of delegated authority: what is routine and what is reserved. Set well, they let an agent move quickly through ordinary work while making sure the larger, riskier operations meet a person before they proceed.

Writing a mandate that works

A good mandate is specific and enforced where the agent cannot reach it. Policy sits below the model, applied by the surrounding system rather than by the agent itself. The agent proposes; the layer beneath accepts or refuses. Limits and allowlists are checked deterministically, and an operation that crosses a threshold becomes an approval request, with the invoice and counterparty attached, instead of a silent failure.

Where the mandate lives

Because the mandate is recorded and versioned, delegated authority stays verifiable over time. Each operation is tied to the policy version it ran under, so a principal can later confirm that an agent acted within the rules that applied. Delegation without a record is a guess; delegation with one is a decision.

In short

  • Limits, approved counterparties and thresholds bound an agent's authority.
  • Policy below the model makes the boundaries unarguable.
  • Over-threshold operations escalate to a human, not a rejection.
  • A versioned mandate keeps delegation verifiable.

See how mandates and controls work on the agent layer.