modbender Analyst Watchdog
An autonomous monitoring agent pattern: watch a system, detect changes, write findings, alert when thresholds are crossed. Runs on a schedule (LaunchAgent or cron) without human intervention.
Pattern
API Endpoint → Poll → Detect Changes → Write FINDINGS.md
→ Write OUTBOX.md (for orchestrator)
→ Write ALERT_TELEGRAM.md (urgent)
What It Monitors
Configurable, but the reference implementation tracks:
- Model evaluation scores hitting milestones (n=50, 100, 150, 200)
- Promotion events (model proven equivalent to cloud baseline)
- Score anomalies (sudden drops or improvements)
- System health degradation
Alert Tiers
| File | Urgency | Who Reads It |
|---|---|---|
FINDINGS.md |
Low | Background knowledge |
OUTBOX.md |
Medium | Orchestrator on next heartbeat |
ALERT_TELEGRAM.md |
High | Sent immediately, then deleted |
Files
scripts/analyst_agent.py— Reference watchdog implementation