The pattern that pays
Agents earn their keep when a task requires reading several sources, applying rules and producing a structured result that a human approves — supplier quote comparison, ticket triage with remediation, month-end reconciliation. Open-ended chat is the least valuable use.
Design against the failure modes
Hallucinated facts, silent tool failures and runaway loops are engineering problems with engineering answers: retrieval with citations, typed tool contracts, step budgets, and a human approval gate before any write to a system of record.
Evaluate like you test software
Golden datasets per workflow, automatic regression on every prompt or model change, and production sampling with human review. If you cannot measure accuracy, you cannot ship an update safely.
Cost and residency
Route simple steps to small models, cache retrieval, and keep data in-region. A well-designed workflow costs cents per task — and stays inside your compliance boundary.
