What Happens When a Tool Call Fails

An agent calls get_shipping_quote(zip_code, weight_kg). In production you observe three distinct failures over a week:

1. The model calls it with zip_code: "N/A" because the user never gave one.
2. The shipping-provider API returns HTTP 503 for about 90 seconds during a provider incident.
3. The call succeeds but returns a 200-line JSON blob of every available carrier and service tier when the user just wanted a ballpark number.

For each failure, say whether you'd handle it in the harness or surface it to the model as data, and what you'd actually implement.