BEHAVIORAL MONITORING

Sentinel

Real-time behavioral monitoring for AI agents.

Know when agents go off-baseline. Before they cause damage. Per-agent behavioral baselines with statistical anomaly detection.

$npm install @authensor/sentinelCopy
USAGE

Baselines build themselves. You just record.

monitor.ts
import { SentinelEngine } from '@authensor/sentinel';

const sentinel = new SentinelEngine();

// Record agent actions — baselines build automatically
sentinel.record({ agentId: 'agent-1', action: 'file.read', latencyMs: 12 });

// Check for anomalies
const alerts = sentinel.check('agent-1');
if (alerts.length > 0) {
  // Agent behavior has deviated from baseline
  alerts.forEach(a => console.warn(a.type, a.message));
}
DETECTION METHODS

Four layers of anomaly detection.

Statistical baselines that adapt to each agent. No thresholds to hand-tune, no ML models to train.

EWMA

Exponentially Weighted Moving Average

Tracks rolling agent behavior with configurable sensitivity. Recent actions weigh more heavily, so baselines adapt naturally while still catching sudden deviations.

CUSUM

Cumulative Sum

Detects persistent shifts in agent behavior patterns. Small, sustained changes that EWMA might miss accumulate until CUSUM triggers an alert.

Chain Depth Tracking

Delegation depth monitoring

Alerts when agent delegation chains exceed safe depths. Prevents runaway recursion where Agent A calls Agent B calls Agent C into unbounded loops.

Fan-Out Detection

Sub-agent spawn monitoring

Flags when one agent spawns too many sub-agents simultaneously. Catches delegation storms before they cascade into uncontrollable parallel execution.

WHAT IT CATCHES

Anomalies that matter, not noise.

Deny rate spikes

Agent suddenly hitting policy denials at 10x baseline rate

Latency anomalies

Agent taking unusually long per action, possible data exfiltration

Volume surges

Agent making far more requests than its established baseline

Delegation storms

Cascading agent-to-agent calls spiraling beyond safe thresholds

TECHNICAL DETAILS

Built for production. Built to disappear.

0
npm dependencies
<1ms
per-check overhead
2
statistical models
100%
pure TypeScript

Start monitoring agent behavior today.