The reliability checklist

• Parse all model output defensively; turn its mistakes into recoverable signals.
• Guarantee one result per tool call.
• Retry transient failures with capped exponential backoff and jitter; don't retry terminal errors.
• Make side-effecting operations idempotent so retries are safe.
• Persist progress so long work is resumable and input is never lost.
• Classify each dependency as fatal or degradable, and handle accordingly.
• Emit a clean error (not a hang) when a stream fails mid-flight.
• Set timeouts on every external call; consider circuit breakers.
• Cap iterations and detect stuck loops.
• Log and measure failures so you can drive them down.

A reliable agent feels calm: when something underneath it breaks, the user gets a clear message or a degraded-but-useful result, finished work is preserved, and nothing hangs or duplicates. That calm is entirely the product of expecting failure and designing for it.