Explanations, tutorials, project notes, opinions, and field reports. New work appears first. Corrections keep their original publication date.
A small model bill can hide review, correction and investigation work. A fictional refund example shows how to price those costs, check daily specialist slots and decide when to hold expansion.
Track refund attempts, deduplicate model usage and join costs to verified case outcomes to compare agent configurations.
A prompt change can send more refunds to specialists without changing code. Decide its release using tested actions, qualified reviewer time and someone who can stop new admissions.
A refund assistant can give a convincing answer while requesting the wrong amount. Check its proposed action, attempted service requests and stored business outcome before shipping a change.
An agent can propose a refund but it cannot grant itself permission. Check order access before a tool runs and repeat the check in the business service before a refund attempt.
An AI workflow can pass its quality checks while leaving too little time for human review. Use a review worksheet and a launch rule that checks qualified staff, unfinished work and decision deadlines before admitting traffic.
A timed-out write may already be complete. Save its operation identity before sending it and keep uncertain outcomes pending. Recover the original receipt through downstream support rather than sending a new business action.
An agent may retry a tool call after a timeout even though the first call finished. Give each action a stable key and save its result so a retry returns that result instead of creating a second ticket, payment, message, or deployment.
A 192-run Microsoft.Extensions.AI benchmark compares sending 8 or 64 tools on every request with a RequestTools discovery function. At 64 tools, billed tokens drop by about 2.5x. At 8 tools, the extra round trip costs more than it saves.
