Agent Stop Conditions

An agent is not production-ready until it has explicit stop conditions: when to halt, when to escalate to a human, and when to fail loudly instead of retrying silently. If the agent has no encoded stopping rules, stop and design them first — unbounded loops are not "autonomous," they are budget-burning bugs.

This is not observability or a circuit breaker you add later. Stop conditions are part of the agent's contract.

When to run

• Designing a new agent or orchestration flow
• Reviewing an agent before production rollout
• Investigating a runaway agent that consumed excessive tokens or made too many tool calls
• Post-mortem after an incident where "the agent didn't know when to stop"

If you cannot name the agent's stop conditions, the agent is unbounded.

The three failure modes

1. Unbounded loops