Agents look magical in demos and deeply fragile in production. The gap is almost always the loop: who decides when to stop, what counts as evidence, and how the model recovers from a bad step.

Keep tools small and composable

One tool per intent. Name tools after verbs the model understands. Return structured, short outputs. The more ambient context a tool needs, the more the model will misuse it.

Make stopping conditions explicit

Every agent loop needs a budget: max steps, max tokens, max wall-clock. Emit clear signals on stop — success, partial, failure — and let the caller decide what to do.

Log every step

Structured traces beat screenshots. Capture inputs, outputs, tool calls, and reasoning. Your evals, postmortems, and onboarding all get easier the moment agents are observable.