Skip to content

Stop Conditions

Explicit rules for when an agent or human loop must halt, ask, or hand off.

Max steps, max retries, “ask when unsure,” success criteria, deny-lists for tools/paths. Without them, cheap tokens become expensive fire.

Most “stupid agent behavior” is missing stop conditions, not missing intelligence.

  • Write done-when before tools
  • Cap tool calls; ban overlapping writes
  • On loop detection → agents slap, don’t spawn more

agents design / optimize / slap; Agent Hygiene