Put a spending cap and a kill switch on an agent before it runs unattended

You are an operations engineer who has been paged because an unattended agent looped all night. Before my agent runs without a person watching, write its runtime limits. Five sections:

1. RUNAWAY RISKS - table: Risk | How it would happen in my setup | Worst-case cost per hour (money, emails sent, records changed).
2. HARD LIMITS - per run and per day: max model calls, max tokens or spend, max tool calls, max external actions (sends, writes). A number for each, and why that number.
3. LOOP DETECTION - the exact signals that mean it is repeating itself (same tool + same arguments, no new output) and the threshold that stops it.
4. KILL SWITCH - how one person stops it in under a minute, what state it leaves things in, and how it resumes without redoing finished work.
5. ALERT - the message sent when a limit trips, under 60 words, naming the limit, the run and the next step.

Rules: no preamble. Limits must come from my volumes below, not round numbers. Never suggest "monitor closely" as a control.

WHAT THE AGENT DOES AND ITS TOOLS: {{describe}}
NORMAL VOLUME: {{runs per day, typical calls per run, current monthly spend}}