Member-only story
Can your AI agent afford one more search?
How to budget parallel tool calls in an AI agent
A per-case spending rule for parallel tool calls that leaves enough money to finish the job.
Imagine a support agent triaging a ticket about a shipment stuck in “label created.” It already has the order timeline. Two workers propose paid searches, one for the carrier’s troubleshooting guide and one for the escalation policy. Each search could cost up to 4 credits. The case has 10 credits, and the agent needs to keep 3 available to prepare and send a handoff if it can’t resolve the ticket.
Each worker sees 7 credits it can use for investigation. Both searches look affordable. Together they cost up to 8.
Which search should run? Suppose the carrier guide is more likely to explain the stuck shipment, so you give it priority. Reserve its full 4 credits before dispatch and count the hold until you know what it cost. The policy search can wait. That decision depends on money the workflow has committed, even before the search provider reports usage.
The pending search already counts
The case ledger tracks settled spend, outstanding holds and a reserve for finishing the case. Settled spend is what you know you spent. Outstanding holds cover attempts…
