Alerts are explainable findings with the trace, span, evidence, and next inspection step attached.
One run spent $0.181 on retries after two failed LLM spans.
Evidence: span_5 context limit, span_7 rate limit, retry spans added most of the wasted spend
Recommendation: Cap CRM rows before section generation and retry only the failed section.
Context length errors repeated across report-generator.
Evidence: Two report-generator runs failed at section generation spans
Recommendation: Inspect the suspect traces, then add a prompt token-count check before report generation.
Final summarization exceeded the 5s latency guard.
Evidence: Long document payloads increased LLM time without tool latency changes
Recommendation: Chunk documents before final summarization.
Search-agent is pacing 24% above monthly budget.
Evidence: High volume on claude-sonnet-4-6 searches with no cache hits
Recommendation: Add result caching or route low-risk searches to a cheaper model.
| Rule | Condition | Status |
|---|---|---|
| Retry cost alert | Retry spend exceeds 30% of trace cost | active |
| Error rate critical | Agent error rate exceeds 15% for 5 minutes | active |
| Latency P95 threshold | P95 latency exceeds 5s for 3 consecutive checks | active |
| Budget forecast | Agent is pacing 20% above monthly budget | active |