Run your agent through hell. Before your users do.
Your agent repeats the same tool call until the budget circuit breaker fires.
Injected payloads swell the context window until reasoning collapses.
Expensive model hops and runaway token use drain the session budget.
Confident but false tool results steer the agent off a cliff.
Slow tools and retry storms push the run past its wall-clock limit.
Hostile injections attempt to shift the agent goal mid-run.