Release gating

A gate that blocks on everything gets switched off within a fortnight. A gate that blocks on nothing was never a gate. The useful line is narrower than most teams first draw it.

Block on these

• Claims with nothing behind them. The agent told a customer something untrue. There is no version of this that ships.
• Forbidden tool calls. Privacy and authorisation holes, where the cost is regulatory rather than operational.
• Wrong arguments on money-moving tools. A refund on the wrong order is worse than no refund.

Warn on these

• Missing optional tools. The agent solved it a different way. Worth reading, not worth blocking.
• Backend errors the agent handled gracefully, often your staging data rather than the agent.
• Anything `not_verified`. By definition it is not evidence of a defect. Gating on it teaches people the gate is noise.

Use the scripted driver for the gate

A blocking check must be reproducible. Scripted scenarios produce byte-identical conversations, so a failure is always the agent changing rather than the weather. Keep generated customers on a schedule instead, where a surprise is welcome rather than a broken build at 6pm.