The Circuit Breaker Designer

Design automatic stopping conditions for this agent.

Agent: [DESCRIBE]
Runaway scenarios: [LOOPS, RETRIES, COST, VOLUME, DURATION]
Current limits: [DESCRIBE, OR "none"]

Produce:
1. Every runaway mode this agent could enter
2. For each: the signal that detects it, the threshold, and the
   action
3. Hard limits: total tool calls, total tokens, total cost, wall
   clock, iterations
4. Behavioural limits: same tool called N times with the same
   arguments, no progress in N steps, repeated identical errors
5. What to do on trip: stop, escalate, degrade, or pause for
   approval
6. What the agent is told when a breaker trips
7. What the user is told
8. What is logged
9. How a tripped run is resumed, if it should be

Then:
10. Which limits should be per-run, per-user, per-hour, and global?
11. What is the default when a limit is not configured? It should
    never be unlimited.
12. What would a well-behaved run look like against these limits —
    am I setting them so tight that normal work trips them?