Investigator Agent
This page lists every autonomous AI investigation — triggered automatically when the system detects trouble, or manually by you.
Where Investigations Come From
- Automatic: the Sentinel watches scores every 15 minutes and queues an investigation when something significant happens — a cluster turning Critical, a score dropping sharply, an alert storm, or any Tier A/S customer at Warning/Critical.
- Ask-AI box (superadmin): type your own question (e.g., "why did Acme's prod cluster slow down yesterday afternoon?") and it's queued as a top-priority investigation. The pending-tasks banner shows it working.
Reading a Report
Each investigation has:
| Field | What it tells you |
|---|---|
| Severity + Confidence | How bad, and how sure the AI is |
| Verdict | Incident (a real problem), False positive (the trigger fired but the data shows nothing wrong) or Insufficient data (the AI could not read what it needed). The list shows a chip next to the severity for the last two |
| Causes | Up to three candidate causes, each marked confirmed, likely or ruled out, with the data behind it |
| Could Not Check | Checks the AI wanted but could not run, and why: read these before trusting a "ruled out" |
| Recommended Actions | Prioritized next steps |
| Knowledge References | Matching past incidents and known bugs by StarRocks version |
| Summary | Executive-ready 2–3 sentences (safe to paste into a customer update) |
Investigations from before 2026-10-21 have no verdict or causes; they show Root Cause and Evidence instead. A run that fails no longer leaves an "Investigation inconclusive" row behind.
Click a row for the full report, including everything the agent looked at.
Exporting to Google Drive
On an investigation's detail view, Export to Drive (superadmin) creates a Google Doc of the report in the shared reports folder and adds it to the index sheet — useful for sharing outside the app. You get links to the doc and the folder.
Investigations cost real LLM tokens. The KPI tiles show volume and cost; the full breakdown is on the LLM Usage page.