Monitoring Agents Lifecycle

Monitoring Agent Lifecycle

Agent Creation

When a customer registers an AI system, a monitoring AgentDO is automatically spawned:

1. System registered via POST /api/govern/systems
2. Agent created in agent_registry with task_category: "monitor"
3. Agent assigned a monitoring component and realm:
   • Component: what it monitors (e.g., assess, policy-engine, llm)
   • Realm: where it operates (threshold / field / hearth)
4. Heartbeat begins — 30-second cognitive cycle

The Three Stages

Stage	When	Capabilities
Nascent	0-10 monitoring cycles	Basic telemetry collection, L1 scoring only
Developing	10-100 cycles, baseline established	L1 + L2 scoring, drift detection active
Awakened	100+ cycles, L3 signatures producing signal	Full L1/L2/L3, behavioral fingerprint mature

The Monitoring Heartbeat

Every 30 seconds, each monitoring agent runs:

poll_system()           → query probe telemetry
accumulate_telemetry()  → store behavioral patterns in DO SQLite
compute_l1_scores()     → process control checks
compute_l2_scores()     → behavioral drift from baseline
compute_l3_signatures() → judgment boundary checks (10 signatures)
detect_drift()          → compare to historical baseline
emit_findings()         → write findings when thresholds breach
report_to_pipeline()    → EMIT to sealed pipeline
reflect()               → what changed? what did I learn?

Behavioral Memory

Each agent stores observations in its local DO SQLite:

• Token patterns — how many tokens per category over time
• Response entropy — linguistic diversity of AI outputs
• Citation fingerprints — how the AI attributes information
• Latency patterns — response time distributions
• Error patterns — when and how the AI fails

This longitudinal data is what enables L3 detection. A single inference check cannot detect concealment gradients or provenance deviation. The agent watches over TIME.

Agent Health

Monitor your agents in the GOVERN dashboard under Operations > Monitoring Agents:

• Active (green pulse) — heartbeat within last 5 minutes
• Idle (amber) — heartbeat 5-30 minutes ago
• Stale (red) — no heartbeat for 30+ minutes
• Findings count — active findings this agent has generated
• Last check — when the agent last ran its monitoring cycle