Agent Fallback

The mechanism by which an agent yields to a human or alternate path when it cannot proceed safely or confidently.

Fallback is what the agent does when it shouldn’t continue: hit a confidence threshold, repeated a failing action, exceeded a budget, encountered an unrecognized situation. The fallback path is usually escalation to a human, but it can also be a deterministic backup script, a different agent, or a graceful refusal.

Designing fallbacks well is operational hygiene. Bad fallback design produces either silent failures (agent gave up but reported success) or panic escalations (every uncertainty hits the human queue). Good fallback design balances autonomy and safety — the agent handles 95% of cases independently, escalates 5% with a clear summary of what it tried and why it stopped.