Success rate dropped

success_rate < 95% over 5m

Latency p95 elevated

p95_latency > 5s over 10m

Token spend spike

spend_rate > $200/hr

Tool error rate

tool_errors / runs > 2%

Timeout cluster

timeouts > 10 in 5m

Eval regression

eval_score < baseline − 0.05

Loop detection

same_tool_calls > 8 in run

Budget at 90%

monthly_spend / budget > 0.9