An execution layer that chooses among many providers is, in effect, a small marketplace. Many instructions arrive at once; several paths are available for each; costs and speeds differ. Making that fair and stable is an engineering problem with familiar tools: rate limits, backpressure and transparent selection.

Why routing needs rules

SUPA's AI routing evaluates every available path — hub participants, card networks, local instant payments, SWIFT, stablecoins, agent protocols — by cost, settlement time, reliability and regulatory fit. Left unmanaged, such a system rewards whoever shouts loudest. Rules make the selection predictable and defensible.

Backpressure over overload

When demand exceeds capacity, the wrong response is to accept everything and degrade silently. Backpressure signals demand early: instructions queue, throughput is shaped, and callers receive an honest indication that the system is busy rather than a timeout after the fact. Idempotency keys mean a client can retry safely while the queue drains. Timeouts are not a form of backpressure; they hide the problem from the caller without relieving the system.

  • shape throughput at the edge, not deep inside
  • queue rather than drop; surface queue state to callers
  • let clients retry with the same idempotency key
  • separate capacity limits from policy limits

Fairness is a design decision

Fairness in an auction is not automatic. Without deliberate rules, participants with more traffic or faster retries dominate. Making selection criteria explicit — and applying them consistently regardless of who is calling — is what keeps the marketplace credible. Because the remainder of netted flow is what actually leaves, selection affects real cost, so the rules should be visible. Where a market sets its rules silently, participants cannot predict outcomes or appeal them. Documented criteria let a provider understand why a given path was chosen, which is also what makes the system auditable.

Practical ingredients

  • explicit limits per caller and per unit
  • queueing with observable depth
  • idempotent instruction handling
  • documented, consistent selection criteria

In short

  • Treat execution as a marketplace that needs explicit rules.
  • Prefer backpressure and queueing to silent degradation.
  • Let idempotency make retries safe during congestion.
  • Make selection criteria explicit so fairness is verifiable.

Read how routing and the order book fit together: protocol.