Silent failures kill agents. Loud ones just wake you up.

The worst agent failure mode is not a crash. It is a quiet success report for work that never happened. Our voice agents were marking calls complete when the booking never landed, callers hung up mid-flow, calendars rejected the slot, tools returned soft errors that the agent treated as done. The dashboard stayed green. Customers did not.

That is the shape of agent risk in production: the model is fluent, the orchestration is optimistic, and the dangerous bugs live in the gap between 'I think I did it' and 'the system of record agrees.' Chat agents fail in a transcript. Voice agents fail out loud, to a stranger, with no undo.

I built an open-source eval harness because manual spot-calls were lying to us. On day one it caught 13 fatal config bugs across 8 production agents, structural lint failures, tool-arg contract violations, KB gaps, persona traps an adversarial caller simulator walks into on purpose. The harness turns every incident into a regression fixture: production-call replay, PHI scrubbed, pinned into CI with an LLM judge under a hard budget ceiling.

The rule I now refuse to negotiate: an agent capability does not exist until there is an eval that fails when it breaks. Vibes-based quality degrades silently, models drift, prompts rot, configs bit-rot, and nobody notices until a customer does. Loud failures page you. Silent ones bill you.

If you are shipping voice or tool-calling agents without regression gates, you are one provider update away from a confident wrong booking. Make the silence impossible.