The demo is lying to you
Every agent works on the happy path. Production starts when users, APIs, latency, budget, and adversarial inputs stop being polite.
Hard limits are not optional. Set max steps, timeout, cost cap, context cap, retry cap, and error threshold. Log the stop reason.
Fail closed when unsure
If the agent cannot verify a high-impact action, it should pause or ask. If cost exceeds budget, stop gracefully. If context pressure rises, compact or hand off state rather than stumbling forward.
Reliability is often the art of stopping early with a useful explanation. Glamorous? no. Necessary? yes.