The gap between an impressive agent demo and an agent that runs your onboarding for six months without supervision is almost entirely engineering — not model quality.
Constrain the action space
An agent that can call any tool will eventually call the wrong one. We give agents a narrow, typed set of actions with explicit preconditions. The model decides which action and with what arguments; the system decides whether that action is allowed right now.
Make every step observable and reversible
Each agent step is logged with its inputs, the model's reasoning, the chosen action and the result. High-impact actions (sending money, emailing a customer, writing to an ERP) pass through a human-in-the-loop queue until confidence and track record justify automation.
Evaluate continuously
We treat agents like any other production service: a golden dataset, regression evals on every prompt or model change, and alerting when accuracy drifts. An agent that silently degrades is worse than no agent.
Done right, agents remove the repetitive 80% and route the ambiguous 20% to a human — which is exactly where the leverage is.