Stop conditions
AI loop stop conditions
A loop without a stop condition is not an operating cycle. It is a token bonfire with a progress bar.
Answer-first definition
A stop condition is the explicit rule that ends an AI loop. Good loops stop on success evidence, budget exhaustion, blocker, no progress, risk boundary, or required human approval.
Six stop conditions
| Stop condition | Meaning | Example |
|---|---|---|
| Success | Verifier passes | Tests exit 0; source table has no unsupported claims. |
| Budget | The loop has spent its cap | Stop after 10 turns, 30 minutes, $5, or 3 failed retries. |
| Blocker | The agent lacks required access or information | Missing API key, ambiguous requirement, locked account. |
| No progress | Repeated attempts do not improve evidence | Same CI failure after two distinct fixes. |
| Risk boundary | The next action is unsafe without review | Production migration, customer email, public post. |
| Human approval | Judgment or commitment is needed | Merge PR, publish page, delete data, spend money. |
Copyable stop clause
Stop when one of these is true:
1. The verifier passes with evidence.
2. You hit 8 turns, 45 minutes, or 3 failed retries.
3. You need credentials, product judgment, or missing requirements.
4. The same failure repeats after two different fixes.
5. The next action would publish, send, delete, deploy, spend money, alter accounts, or create legal/reputational risk.
Return: evidence, diff summary, remaining blocker, and recommended next human decision.